Publish Donto-Qwen3.8-27B V15 adapter, model card, checkpoints, evaluations, sources, and evidence

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+tokenizer.json filter=lfs diff=lfs merge=lfs -text

CHECKSUMS.sha256

@@ -0,0 +1,70 @@
+684888c0ebb17f374298b65ee2807526c066094c701bcc7ebbe1c1095f494fc1  ./.cache/huggingface/.gitignore
+f6572428f6d5e1575e73a1502895a8731f10757dfbb634909c6e154b849af91d  ./.cache/huggingface/CACHEDIR.TAG
+e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855  ./.cache/huggingface/download/LICENSE.lock
+f10121ed978e33b6a028ec65fa64dc4527284e04f699f7f77a0467f529c3ec7d  ./.cache/huggingface/download/LICENSE.metadata
+2c665b32eba1dee799828601543195336a603c797febb25820aab5752a231e18  ./CITATION.cff
+bbedc3fda3305820b977265f01b8619d87570a6739de3a5582c3464840f1e57a  ./LICENSE
+e869b03b051e13fdbebeff829c6a6f7a8e4362356c71bf43a2e9eaef9743bb99  ./NOTICE
+0e5575f8f1e6309d259ad6fca1c24b37bbe19d7209595443b9b5f291510b8d13  ./README.md
+a928e2f4694e2145d7e3c6c7f892bcdc9591a45baa6fa97e380ccb01d495b875  ./adapter_config.json
+40d11bcad20061f8291741c4dd8a3701405902f336c27b60592d62d6b2f7a974  ./adapter_model.safetensors
+62560e8ae9721b4b4562f14d2dac25599cf2a1ea4fb501d7f1b6e925eb6ca711  ./artifacts/TRAINING-COMPLETE.json
+6ed1f5f3e4860e7c38d7eb84d30cf257d3c970bcd754b03a41d3aeb080cdaf4c  ./artifacts/final-validation-metrics.json
+a4ca57bb984b9d54a5a1ebb3da4344e58095bc78d4a841df4854f78b2d4521f8  ./artifacts/lora-target-audit.json
+daa769c6a66465db682d7ce9a977d4541fbc54ac10737e1b5b32ee72c38e5e3d  ./artifacts/run-contract.json
+361cd89313733d55f09f3650c45314b47c4f0d267f00dabc0243a26ec7c27945  ./artifacts/training-manifest.json
+c3cf9e34abf4f9e36c2d72165aa9c132d3e2a725b6c2586aaa3a8af9d7a81041  ./chat_template.jinja
+0a8dbe65cf3f49b8f52a6b8bbd676b69ead955543bdbf2ece2a566e77b0a9ff3  ./docs/FREEZE-RECEIPT-20260825.md
+9c7835a5d54da635bd080fa46ce160b7a06290ec84291f0f23fa6edd912ddf51  ./docs/LAUNCH-RECEIPT.md
+7284a6ad2978bf12381b6494202ad6ec489f7a5d8a69108492f238c8956974be  ./docs/PILOT-LIVE-V11-REPAIR-REJECTION-20260823.md
+fdb7f90d5cf3bbff282e306beb8434ac44bcb5d614684d2f720602c3ad4b64ae  ./docs/PILOT-LIVE-V12-CLEAN-REJECTION-20260823.md
+2c4d0cef6635dbe267932fbafde5c27f84c98936800660d44de407f63d8cda29  ./docs/PILOT-LIVE-V13-FULL746-SEMANTIC-MISMATCH-AUDIT-20260824.md
+1d6ebc4a51aee9649d70f42f742adcdfa25120f911119e9d574272fdc808daae  ./docs/PILOT-LIVE-V14-DENSE-UNIT-REJECTION-20260825.md
+28c8d29bbc76452b26344cf60d36cc0d031c6505d61967b9b90f3d37ec252936  ./docs/PILOT-LIVE-V15-EXHAUSTIVE-UNIT-CORPUS-AUDIT-20260825.md
+09f714c4fa50389c8ea288f022b1dcc6b41a79ddf6635608845bf0c7af789545  ./docs/PILOT-LIVE-V15-EXHAUSTIVE-UNIT-POSTTRAIN-RUNBOOK.md
+c840ed5dc16397892d8ed500edec93b50ca53f085a1a87f1ab76075a46d3bab2  ./docs/QWEN-DONTO-80-PERCENT-GOLD-PLAN-20260822.md
+df9859f741f75a96a958b358a1c08817e5fa6ff4a8fe8eed85a01add501ae7f1  ./docs/UPSTREAM-ADAPTER-README.md
+a21e32e6db21f217af4ad7fc66cbcb3836f0607567bdaab9fcc5c11859e2fd55  ./evaluation/pilot-live-v15-exhaustive-unit-r16-v1-synthetic-gold100-graph-first-v1.json
+a459ed4cad93c641adc89c65a86bb2b1a854628425889ab5d953ad655417822a  ./evaluation/pilot-live-v15-exhaustive-unit-r16-v1-validation-canary1.json
+d9fdbd84546d6f8542f3c1c18643c747316f0238c3f4d0fc7643cf657824d682  ./evaluation/pilot-live-v15-exhaustive-unit-r16-v1-validation-first30.json
+4459693147cf25149ed6241fd935ea0f45406a86dc55110aad51886d228c5408  ./evaluation/pilot-live-v15-exhaustive-unit-r16-v1-validation-full1085.json
+e517d72ceb6443863aaa0a0759b1a021b35dffe25efb41b2fd333b241181b4df  ./examples/structured_client.py
+50c827e62870a6f496a72fb9d8b153775b37c304e2fcd8421b7b49ecb1f0e08c  ./examples/vllm_serve.sh
+358bb8cb878a339f569b1dce8835fe3cf3c8fd1595d9128a35b3c371835c93da  ./requirements.txt
+a873c6e1c16de73d9fc5b964893ff2812307005b87fdc71066cb70fe06d730ed  ./runtime/vllm-conditional/REMAPPING-MANIFEST.json
+1f8abcefc88d6db74239c889fd2246c94b45da3b58c9d68ba8360692b39f06a2  ./runtime/vllm-conditional/adapter_config.json
+28df663ffa70bf541cafff2c9bb60df11e4fe000a2ea9124b346a9cc3eb20d8b  ./runtime/vllm-conditional/adapter_model.safetensors
+c3cf9e34abf4f9e36c2d72165aa9c132d3e2a725b6c2586aaa3a8af9d7a81041  ./runtime/vllm-conditional/chat_template.jinja
+e02cf3c20831eed00a05598f42f324f9112f7c283b9cd0cef43af96e294c93ab  ./schemas/tools.json
+06b9509352d2af50381ab2247e083b80d32d5c0aba91c272ca9ff729b6a0e523  ./tokenizer.json
+66e427c470fe580fe8c7b5725d857af23d8417e37fae62667ec698306a19987b  ./tokenizer_config.json
+a928e2f4694e2145d7e3c6c7f892bcdc9591a45baa6fa97e380ccb01d495b875  ./training/checkpoints/checkpoint-100/adapter_config.json
+ac0fc8a29bdf8bd0de70dd275333cd6cd82e328a161ea86ac44e5e956bb85e67  ./training/checkpoints/checkpoint-100/adapter_model.safetensors
+61f4fb0f34daf13abf304a22501429019f5b65ec145513f2a2ded5685f8e4142  ./training/checkpoints/checkpoint-100/scheduler.pt
+3257b63824c8c43da377913156b18400525c69caea5bca71c671d6eb49b5a75c  ./training/checkpoints/checkpoint-100/trainer_state.json
+a928e2f4694e2145d7e3c6c7f892bcdc9591a45baa6fa97e380ccb01d495b875  ./training/checkpoints/checkpoint-200/adapter_config.json
+8a2dd645a8a3acb7b698820dd0538c355a193825a29513eed313622dcedb16b1  ./training/checkpoints/checkpoint-200/adapter_model.safetensors
+532c3998af11254097768cba2a02356701312926356a1ae080ad6f6e41462901  ./training/checkpoints/checkpoint-200/scheduler.pt
+0f0e0df58888e81b7d798a7e444b81fe3df1e548ee1200c602a94e5a6eb29651  ./training/checkpoints/checkpoint-200/trainer_state.json
+a928e2f4694e2145d7e3c6c7f892bcdc9591a45baa6fa97e380ccb01d495b875  ./training/checkpoints/checkpoint-300/adapter_config.json
+2a65244dee566d1ab4ab27232900310e26d237bf56fabed898b94dde78082d55  ./training/checkpoints/checkpoint-300/adapter_model.safetensors
+841de5b0dcb7b1f112e703b0f661382afd353bc1755625a5460203340f9c49c8  ./training/checkpoints/checkpoint-300/scheduler.pt
+0e1ab6c3974bd751288735034a6a0e376652f6dfec572cff12d2383ca595770f  ./training/checkpoints/checkpoint-300/trainer_state.json
+a928e2f4694e2145d7e3c6c7f892bcdc9591a45baa6fa97e380ccb01d495b875  ./training/checkpoints/checkpoint-400/adapter_config.json
+937f14fcb44105142f2af3c75e8f9b9c3614d3978dc1f228997662dd40777e0b  ./training/checkpoints/checkpoint-400/adapter_model.safetensors
+c42c4df6e3a8d1e7914e0dfeeb99d3defa7a882254b6ae7ab009633a6f88bb3c  ./training/checkpoints/checkpoint-400/scheduler.pt
+3a006e15b2854db0f82937577fd1746b553c11c15d1f03128ad60c977563e72e  ./training/checkpoints/checkpoint-400/trainer_state.json
+a928e2f4694e2145d7e3c6c7f892bcdc9591a45baa6fa97e380ccb01d495b875  ./training/checkpoints/checkpoint-500/adapter_config.json
+ccb510ff3b190867557acdd232bf3f729ba9ea6ee59cd42abdecd937c13f0114  ./training/checkpoints/checkpoint-500/adapter_model.safetensors
+9ff1e247e14de7dd875634e2a89a8a71736f63a4820c8ea62f41d0929f00aab4  ./training/checkpoints/checkpoint-500/scheduler.pt
+c6f7d21f711bb03ecbbf826b81c5cba1295a7f90899e11a7d3f8dc54c390e778  ./training/checkpoints/checkpoint-500/trainer_state.json
+55549bf2ae4398ef1a1ceaacf54cbd0ee2b7527bb416f21868af71b7b130b732  ./training/logs/pilot-live-v15-exhaustive-unit-r16-lr2e5-seed38-v1.log
+7f7985f694a816fc7af105de21801438deb6fedbcfca992feb36e4bbacb64bb9  ./training/logs/pilot-live-v15-exhaustive-unit-r16-v1-bnb-canary.log
+4d479698a708a91ec31a9ac62f194d242b52526704f4874d663960edab10b2bc  ./training/source/audit_rendered_rows.py
+84611c3c6b915099b216245bbe5f6bca9ac0a4de68a2fba257ad43e9ffa527d0  ./training/source/donto-qwen38-eval-corpus.rs
+8a6b2bd79db30000784f46f72377286da85a2cc8e7151f7dc61e4b0f5a3fc8af  ./training/source/donto-qwen38-synthetic-corpus.rs
+f7fd5e06f763f89b8ffd51e615e76d0bfd76154b59f0541f38b4833a92b3e903  ./training/source/donto-qwen38-training-corpus.rs
+2ef9f31ca98ece56ce3732362a1c2917b9375e0e18c260b358ccfcacbfd14572  ./training/source/remap_qwen38_adapter_for_vllm_conditional.py
+236bac7e50741e934fc42e4f6aa512f32b0b14985e908bb60520573c9a81cfee  ./training/source/runpod_train_candidate.sh
+a487c4aea264ee1490ce2eb5fcb7bbce5c3b91351639ddbfdfe55da93487e937  ./training/source/start_runpod_quality_canary.sh
+994d3f888fe3d18f115ac050d59aa68a524ef160dc7f2073a83d4a8568b82d79  ./training/source/train_qwen38_donto_3090.py

CITATION.cff

@@ -0,0 +1,13 @@
+cff-version: 1.2.0
+message: "If you use this adapter, please cite it as below."
+title: "Donto-Qwen3.8-27B Predicate Extractor V15"
+type: software
+authors:
+  - family-names: Davis
+    given-names: Thomas
+repository-code: "https://huggingface.co/ajaxdavis/donto-qwen3.8-27b-predicate-extractor"
+url: "https://huggingface.co/ajaxdavis/donto-qwen3.8-27b-predicate-extractor"
+license: Apache-2.0
+version: "15"
+date-released: "2026-08-25"
+

LICENSE

@@ -0,0 +1,202 @@
+
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright 2026 Alibaba Cloud
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

NOTICE

@@ -0,0 +1,15 @@
+Donto-Qwen3.8-27B Predicate Extractor V15
+
+Copyright 2026 Thomas Davis
+
+This distribution contains a LoRA adapter derived from Qwen/Qwen3.8-27B at
+revision 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0. Qwen3.8-27B is distributed
+under the Apache License 2.0. The upstream base weights are not included here.
+
+The runtime, training, and evaluation stack uses open-source projects including
+Transformers, PEFT, PyTorch, bitsandbytes, Liger Kernel, and vLLM. Their respective
+licences remain applicable.
+
+The linked dataset has separate mixed-provenance licensing terms. Model and code
+licensing does not automatically relicense source excerpts in that dataset.
+

README.md

@@ -0,0 +1,427 @@
+---
+license: apache-2.0
+base_model: Qwen/Qwen3.8-27B
+base_model_relation: adapter
+library_name: peft
+pipeline_tag: text-generation
+language:
+- en
+datasets:
+- ajaxdavis/donto-qwen3.8-27b-predicate-extraction-data
+tags:
+- qwen3.8
+- qwen3_5
+- peft
+- lora
+- qlora
+- information-extraction
+- relation-extraction
+- knowledge-graph
+- structured-output
+- tool-calling
+- donto
+model-index:
+- name: Donto-Qwen3.8-27B Predicate Extractor V15
+  results:
+  - task:
+      type: text-generation
+      name: Structured predicate extraction
+    dataset:
+      type: ajaxdavis/donto-qwen3.8-27b-predicate-extraction-data
+      name: Graph-First Gold 100
+      split: test
+    metrics:
+    - type: exact_recall
+      name: Exact fact recall
+      value: 0.8920944173847883
+    - type: exact_precision
+      name: Exact fact precision
+      value: 0.8944402704733283
+    - type: tool_call_success_rate
+      name: Structured tool-call success
+      value: 1.0
+---
+
+# Donto-Qwen3.8-27B Predicate Extractor
+
+> **An experimental, anchor-free, abundance-oriented predicate-extraction LoRA for short English document chunks.**
+
+This repository contains the selected **V15** adapter for
+[`Qwen/Qwen3.8-27B`](https://huggingface.co/Qwen/Qwen3.8-27B), pinned to base
+revision `1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0`. It turns a general Qwen3.8
+language backbone into a specialist that emits validated Donto fact objects
+instead of conversational prose.
+
+The headline result is **89.21% exact recall and 89.44% exact precision** over a
+frozen, source- and template-disjoint, 100-document graph-first synthetic gold
+suite containing 5,338 target facts. All 2,718 forced structured calls succeeded.
+
+This is a research release, not a claim that open-world extraction is solved.
+Read [What the score does and does not mean](#what-the-score-does-and-does-not-mean)
+and [Limitations](#limitations) before deploying it.
+
+## The ambition
+
+Knowledge systems were designed for a world in which producing structured claims
+was scarce. Generative models invert that constraint: candidate properties and
+relations can be produced in abundance. The new problem is not merely generating
+one summary or one canonical graph. It is holding a large, multi-directional,
+possibly contradictory field of claims without erasing disagreement.
+
+Donto is built for that second problem. It is a contradiction-preserving claim
+substrate: emit free predicates now; preserve separate claims; defer ontology
+alignment, entity resolution, joining, standing, and reconciliation to later
+query-time processes. **Donto-Qwen is the extraction front door.** Its job is to
+read a bounded source unit carefully and produce many compact candidate claims
+that the larger Donto system can hold, compare, cite, align, and re-rank.
+
+The purpose of this adapter is deliberately narrower than general chat:
+
+1. recover atomic subject–predicate–object relations;
+2. preserve attribution, negation, modality, qualification, and disagreement;
+3. represent direct claims separately from derived or hypothetical views;
+4. use free, concise predicates instead of forcing every source into a fixed
+   ontology;
+5. stop cleanly when a bounded extraction task is exhausted;
+6. return only a validated tool call—never prose around hand-parsed JSON.
+
+## What this model is
+
+| Property | Value |
+|---|---|
+| Release | V15 / `pilot-live-v15-exhaustive-unit-r16-v1` |
+| Base | Qwen3.8-27B, exact revision `1d4bf0f...` |
+| Artifact | PEFT LoRA adapter; base weights are not duplicated here |
+| Primary modality | Text; the Qwen vision tower was neither trained nor evaluated |
+| Intended context | Short document chunks; trained and served at 4,096 tokens |
+| Output | Forced `submit_facts` tool call |
+| Fact fields | `s`, `p`, `o`, `c`, `h` |
+| Anchors | Disabled in this release |
+| Thinking | Disabled for training and evaluated serving |
+| Training hardware | One NVIDIA RTX 3090, 24 GB |
+| Training method | 4-bit NF4 QLoRA, BF16 compute, rank 16 |
+| Status | Experimental predicate extractor; not a general assistant |
+
+This adapter changes the base model's **behaviour**, not its underlying factual
+knowledge. It does not intrinsically make Qwen3.8 decode faster. Its efficiency
+advantages come from short inputs, compact structured outputs, shared prompt
+prefixes, fewer retries, and greater document concurrency.
+
+## Output contract
+
+Every call must produce the `submit_facts` tool with this shape:
+
+```json
+{
+  "facts": [
+    {
+      "s": "ex:review-panel",
+      "p": "postponed",
+      "o": "ex:deployment",
+      "c": 0.99,
+      "h": false
+    },
+    {
+      "s": "ex:postponement-wasted-money",
+      "p": "contrastsWith",
+      "o": "ex:postponement-protected-staff",
+      "c": 0.92,
+      "h": true
+    }
+  ],
+  "more_supported_facts": false
+}
+```
+
+- `s`: subject IRI, normally a stable `ex:kebab-case` identifier.
+- `p`: concise camelCase free predicate, with standard RDF predicates allowed.
+- `o`: entity IRI or string, numeric, or Boolean literal.
+- `c`: model confidence in `[0, 1]`; this is **not calibrated probability**.
+- `h`: `false` for direct/source-licensed claims; `true` for a derived,
+  interpretive, hypothetical, or otherwise non-direct view.
+- `more_supported_facts`: continuation signal for the current bounded task.
+
+The exact OpenAI-compatible tool declaration is in [`schemas/tools.json`](schemas/tools.json).
+
+## Extraction architecture used for the reported scores
+
+The benchmark did not ask one vague prompt to "extract everything." Each document
+was traversed through explicit bounded tasks:
+
+| Stage | Purpose |
+|---|---|
+| Metadata | Document, creator, date, and topic relations |
+| D0 | Atomic propositions directly expressed in a source unit |
+| Local residual | Recover a deliberately omitted supported fact, then prove exhaustion |
+| D1 unit | Licensed semantic views and local relational expansion |
+| D2 | Discourse-level, contrastive, causal, or interpretive relations (`h:true`) |
+| Structural batches | Relations over the dynamically discovered entity inventory |
+| Global residual | Final bounded coverage audit |
+
+This decomposition is important. A single generic extraction prompt is not
+expected to reproduce the published multi-stage score.
+
+## Results
+
+### Headline evaluations
+
+| Evaluation | Documents | Calls | Gold facts | Exact recall | Exact precision | Mean document recall | 95% bootstrap recall floor | Structured success |
+|---|---:|---:|---:|---:|---:|---:|---:|---:|
+| Source-disjoint V15 validation | 40 | 1,085 | 2,140 | **93.32%** | **93.89%** | 93.36% | 91.74% | 100% |
+| Frozen graph-first Gold 100 | 100 | 2,718 | 5,338 | **89.21%** | **89.44%** | 89.50% | **87.82%** | 100% |
+
+Gold 100 also produced:
+
+- 4,762 exact matches from 5,338 targets;
+- 5,324 valid predicted facts;
+- zero failed calls;
+- zero malformed fact objects;
+- zero duplicate exact items;
+- two reflexive entity errors.
+
+### Gold 100 by task
+
+| Task | Calls | Targets | Exact recall | Exact precision |
+|---|---:|---:|---:|---:|
+| D0 direct extraction | 428 | 1,751 | **75.10%** | **75.14%** |
+| D1 unit views | 766 | 645 | 95.35% | 97.77% |
+| D2 discourse views | 212 | 446 | 93.72% | 94.78% |
+| Metadata | 100 | 361 | 95.01% | 95.28% |
+| Global residual | 200 | 300 | 96.33% | 96.01% |
+| Local residual | 590 | 162 | 97.53% | 98.14% |
+| Structural batches | 322 | 1,673 | 97.07% | 96.55% |
+
+The overall release clears the project's predeclared 80% aggregate-gold target,
+but the direct D0 slice does not. That distinction is intentional and public.
+The weakest blind families were long wetlands argument/procurement documents;
+the strongest were bounded incidents, metadata, residual stopping, and structural
+relations.
+
+Full row-level outputs—not summaries alone—are published under [`evaluation/`](evaluation/).
+
+## What the score does and does not mean
+
+The metric uses one-to-one **exact keys** over `(s, p, o, h)`. A supported free
+predicate synonym or stable entity alias receives no credit, which can make the
+score conservative. Conversely, the gold documents are programmatically generated
+from explicit graphs and templates. They measure controlled compositional transfer,
+not the full ambiguity, OCR damage, genre diversity, or factual uncertainty of
+unseen real-world documents.
+
+The 100-document suite was generated after V15 training, then frozen and hashed
+before the first V15 inference. It has:
+
+- 100 unique document hashes;
+- 73 parent templates;
+- zero source-hash overlap with the 240 V15 training/validation documents;
+- zero parent-template overlap with the 102 training/validation templates;
+- 45–59 unique gold facts per document (mean 52.88).
+
+The run used one seed. It is not a multi-seed reliability estimate and is not a
+human-blind real-document benchmark. The prior real `llm_nerd` fixture was heavily
+used for development diagnosis and is therefore not presented as blind evidence.
+
+## Training data
+
+All distributable training, validation, validator, gold, manifest, and audit files
+are published at
+[`ajaxdavis/donto-qwen3.8-27b-predicate-extraction-data`](https://huggingface.co/datasets/ajaxdavis/donto-qwen3.8-27b-predicate-extraction-data).
+
+Canonical V15 training mixture:
+
+- 4,255 training rows;
+- 9,013 target facts;
+- 200 source families;
+- 4,123 synthetic task rows from 152 generated documents;
+- 132 reviewed-real rows from 48 source families;
+- 1,085 validation rows with 2,140 facts from 40 source- and template-disjoint
+  synthetic documents;
+- no training/validation source-family overlap;
+- no training/validation parent-template overlap;
+- every rendered row below the 4,096-token contract.
+
+The dataset card separates synthetic and reviewed-real provenance and explains its
+mixed licensing status.
+
+## Training recipe
+
+| Parameter | Value |
+|---|---|
+| Base revision | `1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0` |
+| Seed | 38 |
+| Epochs / optimizer steps | 1 / 532 |
+| LoRA rank / alpha / dropout | 16 / 32 / 0.05 |
+| Trainable parameters | 116,727,808 |
+| Learning rate | `2e-5`, cosine decay |
+| Warmup | 5% / 27 steps |
+| Microbatch / accumulation | 1 / 8 |
+| Maximum length | 4,096 |
+| Quantization | bitsandbytes NF4, double quantization |
+| Compute | BF16; TF32 enabled |
+| Optimizer | paged AdamW 8-bit |
+| Kernel | Liger enabled |
+| Gradient checkpointing | Enabled, non-reentrant |
+| Train runtime | 20,860 seconds including epoch evaluation |
+| Final explicit eval loss | `0.0010853700805455446` |
+
+### Hybrid-layer LoRA coverage
+
+Qwen3.8-27B contains 48 Gated DeltaNet layers and 16 full-attention layers. A
+generic `q_proj`/`v_proj` LoRA recipe would miss most recurrent projections.
+This adapter explicitly targeted **496 language modules / 992 LoRA tensors** and
+no vision module:
+
+- 48 each: `linear_attn.in_proj_qkv`, `in_proj_z`, `in_proj_a`, `in_proj_b`,
+  `out_proj`;
+- 16 each: `self_attn.q_proj`, `k_proj`, `v_proj`, `o_proj`;
+- 64 each: `mlp.gate_proj`, `up_proj`, `down_proj`.
+
+The complete name-level audit is [`artifacts/lora-target-audit.json`](artifacts/lora-target-audit.json).
+
+## Serving with vLLM on one RTX 3090
+
+The physically tested path used vLLM 0.27.1, Transformers 5.15.0,
+bitsandbytes 0.50.1, a 4,096-token limit, four concurrent sequences, and prefix
+caching. vLLM required the tensor-preserving conditional-key remap under
+[`runtime/vllm-conditional/`](runtime/vllm-conditional/); the root adapter remains
+the canonical native PEFT artifact.
+
+```bash
+hf download ajaxdavis/donto-qwen3.8-27b-predicate-extractor \
+  --local-dir ./donto-qwen
+
+vllm serve Qwen/Qwen3.8-27B \
+  --served-model-name qwen3.8-27b \
+  --language-model-only \
+  --quantization bitsandbytes \
+  --load-format bitsandbytes \
+  --dtype bfloat16 \
+  --max-model-len 4096 \
+  --max-num-seqs 4 \
+  --max-num-batched-tokens 2048 \
+  --enable-prefix-caching \
+  --generation-config vllm \
+  --reasoning-parser qwen3 \
+  --enable-auto-tool-choice \
+  --tool-call-parser qwen3_coder \
+  --enable-lora \
+  --lora-modules donto-qwen=./donto-qwen/runtime/vllm-conditional \
+  --max-loras 1 \
+  --max-lora-rank 16
+```
+
+The exact tested launcher includes additional GPU-specific cache and attention
+flags and is preserved at [`training/source/start_runpod_quality_canary.sh`](training/source/start_runpod_quality_canary.sh).
+
+Then call the forced tool and validate its arguments with a real schema. A complete
+Pydantic/OpenAI-compatible example is in [`examples/structured_client.py`](examples/structured_client.py).
+
+## Native PEFT use
+
+The root `adapter_model.safetensors` is the canonical PEFT adapter. Load it over
+the exact pinned base revision with current Transformers and PEFT. The training
+script in [`training/source/train_qwen38_donto_3090.py`](training/source/train_qwen38_donto_3090.py)
+contains the memory-safe text-only 3090 loader and exact QLoRA configuration.
+
+Do not attach the root native adapter directly to vLLM if its Qwen3.8 key resolver
+expects the conditional language-model root; use the verified remap supplied in
+`runtime/vllm-conditional`.
+
+## Efficiency profile
+
+- Adapter weights are about 467 MB; the base model remains the dominant memory
+  and compute cost.
+- The adapter does not improve raw decode tokens/second by itself.
+- Short chunks reduce prefill and KV-cache costs.
+- Prefix caching reuses the large common tool/prompt prefix across documents.
+- The anchor-free five-field format reduces output tokens per useful fact.
+- Four concurrent sequences fit in the physically tested 24 GB 3090 serving path.
+- Gold 100 completed 2,718 structured calls in 8,511 seconds at concurrency four,
+  about 19.2 bounded decisions per minute. This is not a generic tokens/second
+  benchmark.
+
+## Limitations
+
+1. **Direct extraction is the weakest blind slice.** Gold-100 D0 exact recall and
+   precision are both about 75%.
+2. **Synthetic-domain bias.** The main quantitative evidence comes from generated
+   graph/template documents, not a large human-authored blind corpus.
+3. **One training seed.** The recipe has not been repeated across multiple seeds.
+4. **Two reflexive errors occurred** in Gold 100. Role direction and reflexivity
+   still require downstream validation.
+5. **No evidence anchors.** This release deliberately omits source spans. A separate
+   citer or human review is required when provenance is important.
+6. **Free predicates are not ontology alignment.** Predicate and entity aliases may
+   vary; alignment belongs downstream.
+7. **Confidence is uncalibrated.** Do not interpret `c` as a probability of truth.
+8. **`h:true` is not established fact.** It marks derived or interpretive output.
+9. **Not a contradiction resolver.** It may emit incompatible claims by design;
+   Donto preserves and later ranks them.
+10. **Not for unattended high-stakes use.** Legal, medical, financial, historical,
+    or personal claims require source review.
+11. **Not a general assistant.** Fine-tuning trades conversational versatility for
+    structured extraction behaviour.
+12. **Multimodal behaviour is unevaluated.** No vision module was adapted.
+
+## Development history and negative results
+
+This release is the fifteenth curriculum iteration, trained fresh from the pinned
+stock base. Rejected adapters were never reused as weight bases. The repository
+preserves rejection reports for V11–V14 because the failures are informative:
+
+- V11 learned structured calls but missed its precision gate and had semantic
+  contamination in one generator family.
+- V12 completed finitely but was too terse and failed real-document coverage.
+- V13 reached strong source-disjoint exact metrics but retained role/scope errors.
+- V14 nearly memorized its supervised contract (about 99% exact) yet transferred
+  poorly to the real development document because its curriculum rewarded short,
+  clean pages.
+- V15 started fresh and taught dense D0 pages, nonempty recovery passes, explicit
+  exhaustion, larger local D1 views, and broader graph families.
+
+See [`docs/`](docs/) for the plan, audits, launch receipt, runbook, and complete
+negative-result reports.
+
+## Reproducibility and integrity
+
+- Native final adapter SHA-256:
+  `40d11bcad20061f8291741c4dd8a3701405902f336c27b60592d62d6b2f7a974`
+- vLLM remapped adapter SHA-256:
+  `28df663ffa70bf541cafff2c9bb60df11e4fe000a2ea9124b346a9cc3eb20d8b`
+- Training corpus SHA-256:
+  `1b3630d00d11476c7f1e941a60a442a5eb5eb57dd882b7d188ef9b85d461e738`
+- V15 validation SHA-256:
+  `c360e8296e5ec7973f2fab8416d0b30e2aa850fe6b2cc618bb3fa52894d818b2`
+- Gold-100 dataset SHA-256:
+  `6f16c87d06c2a0b850321b8750e54a63364937cfa2823f34f698f6c00f160f1c`
+- Gold-100 evaluation SHA-256:
+  `a21e32e6db21f217af4ad7fc66cbcb3836f0607567bdaab9fcc5c11859e2fd55`
+
+The repository includes intermediate adapter checkpoints 100–500, trainer states,
+the exact trainer/evaluator/generator sources, training and serving logs, run
+contracts, target-module audit, and full row-level evaluation outputs. Unsafe and
+unnecessary optimizer/RNG pickle states are not part of the public release.
+
+## Licence and acknowledgements
+
+The adapter and code are released under Apache-2.0, consistent with the pinned
+Qwen3.8-27B base. Dataset licensing is documented separately because the corpus
+contains both generated material and a small reviewed-real research tier.
+
+Built on Qwen3.8-27B, PEFT, Transformers, bitsandbytes, Liger Kernel, and vLLM.
+The practical RTX 3090 serving work was informed by
+[`syv-ai/qwen38-27b-rtx3090`](https://github.com/syv-ai/qwen38-27b-rtx3090).
+
+## Citation
+
+```bibtex
+@misc{donto_qwen38_v15_2026,
+  author       = {Davis, Thomas},
+  title        = {Donto-Qwen3.8-27B Predicate Extractor V15},
+  year         = {2026},
+  publisher    = {Hugging Face},
+  howpublished = {\url{https://huggingface.co/ajaxdavis/donto-qwen3.8-27b-predicate-extractor}},
+  note         = {Experimental LoRA adapter for structured Donto predicate extraction}
+}
+```
+

adapter_config.json

@@ -0,0 +1,42 @@
+{
+  "alora_invocation_tokens": null,
+  "alpha_pattern": {},
+  "arrow_config": null,
+  "auto_mapping": null,
+  "base_model_name_or_path": "Qwen/Qwen3.8-27B",
+  "bias": "none",
+  "corda_config": null,
+  "ensure_weight_tying": false,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 32,
+  "lora_bias": false,
+  "lora_dropout": 0.05,
+  "lora_ga_config": null,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "monteclora_config": null,
+  "peft_type": "LORA",
+  "peft_version": "0.20.0",
+  "qalora_group_size": 16,
+  "r": 16,
+  "rank_pattern": {},
+  "revision": "1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0",
+  "target_modules": ".*(?:language_model\\.)?layers\\.\\d+\\.(?:self_attn\\.(?:q_proj|k_proj|v_proj|o_proj)|linear_attn\\.(?:in_proj_qkv|in_proj_z|in_proj_a|in_proj_b|out_proj)|mlp\\.(?:gate_proj|up_proj|down_proj))$",
+  "target_parameters": null,
+  "task_type": "CAUSAL_LM",
+  "trainable_token_indices": null,
+  "use_bdlora": null,
+  "use_dora": false,
+  "use_qalora": false,
+  "use_rslora": false,
+  "velora_config": null
+}

adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:40d11bcad20061f8291741c4dd8a3701405902f336c27b60592d62d6b2f7a974
+size 467047680

artifacts/CHECKSUMS.sha256

@@ -0,0 +1,12 @@
+62560e8ae9721b4b4562f14d2dac25599cf2a1ea4fb501d7f1b6e925eb6ca711  TRAINING-COMPLETE.json
+6ed1f5f3e4860e7c38d7eb84d30cf257d3c970bcd754b03a41d3aeb080cdaf4c  final-validation-metrics.json
+df9859f741f75a96a958b358a1c08817e5fa6ff4a8fe8eed85a01add501ae7f1  final/README.md
+a928e2f4694e2145d7e3c6c7f892bcdc9591a45baa6fa97e380ccb01d495b875  final/adapter_config.json
+40d11bcad20061f8291741c4dd8a3701405902f336c27b60592d62d6b2f7a974  final/adapter_model.safetensors
+c3cf9e34abf4f9e36c2d72165aa9c132d3e2a725b6c2586aaa3a8af9d7a81041  final/chat_template.jinja
+06b9509352d2af50381ab2247e083b80d32d5c0aba91c272ca9ff729b6a0e523  final/tokenizer.json
+66e427c470fe580fe8c7b5725d857af23d8417e37fae62667ec698306a19987b  final/tokenizer_config.json
+175111135c9d40b63054d1ef321345f80de8c4d21717241624ae93bced357d6c  final/training_args.bin
+a4ca57bb984b9d54a5a1ebb3da4344e58095bc78d4a841df4854f78b2d4521f8  lora-target-audit.json
+daa769c6a66465db682d7ce9a977d4541fbc54ac10737e1b5b32ee72c38e5e3d  run-contract.json
+361cd89313733d55f09f3650c45314b47c4f0d267f00dabc0243a26ec7c27945  training-manifest.json

artifacts/TRAINING-COMPLETE.json

@@ -0,0 +1,7 @@
+{
+  "blind_holdout_opened": false,
+  "final_adapter_config_sha256": "a928e2f4694e2145d7e3c6c7f892bcdc9591a45baa6fa97e380ccb01d495b875",
+  "final_adapter_sha256": "40d11bcad20061f8291741c4dd8a3701405902f336c27b60592d62d6b2f7a974",
+  "metrics_sha256": "6ed1f5f3e4860e7c38d7eb84d30cf257d3c970bcd754b03a41d3aeb080cdaf4c",
+  "status": "training_complete_not_yet_accepted"
+}

artifacts/final-validation-metrics.json

@@ -0,0 +1,7 @@
+{
+  "epoch": 1.0,
+  "eval_loss": 0.0010853700805455446,
+  "eval_runtime": 1556.9529,
+  "eval_samples_per_second": 0.697,
+  "eval_steps_per_second": 0.697
+}

artifacts/run-contract.json

@@ -0,0 +1,36 @@
+{
+  "base_model": "/workspace/models/Qwen3.8-27B-BF16",
+  "base_revision_requested": "1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0",
+  "contract": "donto-qwen38-bounded-task-no-thinking-completion-only-v7",
+  "dataset_audit_sha256": "7cd156217b16636a2c918ddd737dd33b0ee2a712e773dab0984a0dcd91d52c7d",
+  "dataset_manifest_sha256": "26e3376dc6b99f8b1392a868dd0365a65ae0f1e4e20721cf994ad2f46d1a3231",
+  "epochs": 1.0,
+  "first_train_prompt_token_ids_sha256": "5ea6c0d6c390fa133f0df6940952e4d7c581839d337769d4df84ac5b301d2a66",
+  "first_train_prompt_tokens": 1134,
+  "first_validation_prompt_token_ids_sha256": "55db291b7138c75b1cb0300e15536b8da8ccc2d8b668630c628fa5ae115460d9",
+  "first_validation_prompt_tokens": 1133,
+  "gradient_accumulation": 8,
+  "learning_rate": 2e-05,
+  "liger_kernel": true,
+  "lora_rank": 16,
+  "lora_use_rslora": false,
+  "max_len": 4096,
+  "max_steps": -1,
+  "run_id": "pilot-live-v15-exhaustive-unit-r16-lr2e5-seed38-v1",
+  "save_steps": 100,
+  "seed": 38,
+  "smoke_longest": false,
+  "tool_name": "submit_facts",
+  "tools_sha256": "e02cf3c20831eed00a05598f42f324f9112f7c283b9cd0cef43af96e294c93ab",
+  "tools_transport_sha256": "8945d9bfd5e68ea50537c4cae5476fcfa9efd4e99ab80c9b71bc4bfcbd5826b1",
+  "train_files": [
+    {
+      "name": "train.jsonl",
+      "sha256": "1b3630d00d11476c7f1e941a60a442a5eb5eb57dd882b7d188ef9b85d461e738"
+    }
+  ],
+  "train_sha256": "1b3630d00d11476c7f1e941a60a442a5eb5eb57dd882b7d188ef9b85d461e738",
+  "trainer_source_sha256": "994d3f888fe3d18f115ac050d59aa68a524ef160dc7f2073a83d4a8568b82d79",
+  "validation_sha256": "c360e8296e5ec7973f2fab8416d0b30e2aa850fe6b2cc618bb3fa52894d818b2",
+  "warmup_ratio": 0.05
+}

artifacts/training-manifest.json

@@ -0,0 +1,324 @@
+{
+  "base_model": "/workspace/models/Qwen3.8-27B-BF16",
+  "base_revision_requested": "1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0",
+  "base_revision_resolved": "1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0",
+  "blind_holdout_opened": false,
+  "contract": "donto-qwen38-bounded-task-no-thinking-completion-only-v7",
+  "dataset_audit_sha256": "7cd156217b16636a2c918ddd737dd33b0ee2a712e773dab0984a0dcd91d52c7d",
+  "dataset_manifest_sha256": "26e3376dc6b99f8b1392a868dd0365a65ae0f1e4e20721cf994ad2f46d1a3231",
+  "epochs": 1.0,
+  "first_train_prompt_token_ids_sha256": "5ea6c0d6c390fa133f0df6940952e4d7c581839d337769d4df84ac5b301d2a66",
+  "first_train_prompt_tokens": 1134,
+  "first_validation_prompt_token_ids_sha256": "55db291b7138c75b1cb0300e15536b8da8ccc2d8b668630c628fa5ae115460d9",
+  "first_validation_prompt_tokens": 1133,
+  "gradient_accumulation": 8,
+  "host": {
+    "bf16_supported": true,
+    "cuda_version": "12.8",
+    "gpu_name": "NVIDIA GeForce RTX 3090",
+    "gpu_total_gib": 23.558,
+    "output_free_gib_before_training": 308835.961,
+    "qwen3_5_recurrent_kernels_available": true
+  },
+  "learning_rate": 2e-05,
+  "liger_kernel": true,
+  "lora_rank": 16,
+  "lora_target_audit": {
+    "artifact": "lora-target-audit.json",
+    "targeted_module_count": 496,
+    "targeted_parameter_count": 992
+  },
+  "lora_use_rslora": false,
+  "max_len": 4096,
+  "max_steps": -1,
+  "memory_safe_3090_prepare": {
+    "frozen_embedding_dtype": "torch.bfloat16",
+    "upcast_norm_parameter_count": 209,
+    "upcast_norm_parameter_names": [
+      "model.layers.0.linear_attn.norm.weight",
+      "model.layers.0.input_layernorm.weight",
+      "model.layers.0.post_attention_layernorm.weight",
+      "model.layers.1.linear_attn.norm.weight",
+      "model.layers.1.input_layernorm.weight",
+      "model.layers.1.post_attention_layernorm.weight",
+      "model.layers.2.linear_attn.norm.weight",
+      "model.layers.2.input_layernorm.weight",
+      "model.layers.2.post_attention_layernorm.weight",
+      "model.layers.3.self_attn.q_norm.weight",
+      "model.layers.3.self_attn.k_norm.weight",
+      "model.layers.3.input_layernorm.weight",
+      "model.layers.3.post_attention_layernorm.weight",
+      "model.layers.4.linear_attn.norm.weight",
+      "model.layers.4.input_layernorm.weight",
+      "model.layers.4.post_attention_layernorm.weight",
+      "model.layers.5.linear_attn.norm.weight",
+      "model.layers.5.input_layernorm.weight",
+      "model.layers.5.post_attention_layernorm.weight",
+      "model.layers.6.linear_attn.norm.weight",
+      "model.layers.6.input_layernorm.weight",
+      "model.layers.6.post_attention_layernorm.weight",
+      "model.layers.7.self_attn.q_norm.weight",
+      "model.layers.7.self_attn.k_norm.weight",
+      "model.layers.7.input_layernorm.weight",
+      "model.layers.7.post_attention_layernorm.weight",
+      "model.layers.8.linear_attn.norm.weight",
+      "model.layers.8.input_layernorm.weight",
+      "model.layers.8.post_attention_layernorm.weight",
+      "model.layers.9.linear_attn.norm.weight",
+      "model.layers.9.input_layernorm.weight",
+      "model.layers.9.post_attention_layernorm.weight",
+      "model.layers.10.linear_attn.norm.weight",
+      "model.layers.10.input_layernorm.weight",
+      "model.layers.10.post_attention_layernorm.weight",
+      "model.layers.11.self_attn.q_norm.weight",
+      "model.layers.11.self_attn.k_norm.weight",
+      "model.layers.11.input_layernorm.weight",
+      "model.layers.11.post_attention_layernorm.weight",
+      "model.layers.12.linear_attn.norm.weight",
+      "model.layers.12.input_layernorm.weight",
+      "model.layers.12.post_attention_layernorm.weight",
+      "model.layers.13.linear_attn.norm.weight",
+      "model.layers.13.input_layernorm.weight",
+      "model.layers.13.post_attention_layernorm.weight",
+      "model.layers.14.linear_attn.norm.weight",
+      "model.layers.14.input_layernorm.weight",
+      "model.layers.14.post_attention_layernorm.weight",
+      "model.layers.15.self_attn.q_norm.weight",
+      "model.layers.15.self_attn.k_norm.weight",
+      "model.layers.15.input_layernorm.weight",
+      "model.layers.15.post_attention_layernorm.weight",
+      "model.layers.16.linear_attn.norm.weight",
+      "model.layers.16.input_layernorm.weight",
+      "model.layers.16.post_attention_layernorm.weight",
+      "model.layers.17.linear_attn.norm.weight",
+      "model.layers.17.input_layernorm.weight",
+      "model.layers.17.post_attention_layernorm.weight",
+      "model.layers.18.linear_attn.norm.weight",
+      "model.layers.18.input_layernorm.weight",
+      "model.layers.18.post_attention_layernorm.weight",
+      "model.layers.19.self_attn.q_norm.weight",
+      "model.layers.19.self_attn.k_norm.weight",
+      "model.layers.19.input_layernorm.weight",
+      "model.layers.19.post_attention_layernorm.weight",
+      "model.layers.20.linear_attn.norm.weight",
+      "model.layers.20.input_layernorm.weight",
+      "model.layers.20.post_attention_layernorm.weight",
+      "model.layers.21.linear_attn.norm.weight",
+      "model.layers.21.input_layernorm.weight",
+      "model.layers.21.post_attention_layernorm.weight",
+      "model.layers.22.linear_attn.norm.weight",
+      "model.layers.22.input_layernorm.weight",
+      "model.layers.22.post_attention_layernorm.weight",
+      "model.layers.23.self_attn.q_norm.weight",
+      "model.layers.23.self_attn.k_norm.weight",
+      "model.layers.23.input_layernorm.weight",
+      "model.layers.23.post_attention_layernorm.weight",
+      "model.layers.24.linear_attn.norm.weight",
+      "model.layers.24.input_layernorm.weight",
+      "model.layers.24.post_attention_layernorm.weight",
+      "model.layers.25.linear_attn.norm.weight",
+      "model.layers.25.input_layernorm.weight",
+      "model.layers.25.post_attention_layernorm.weight",
+      "model.layers.26.linear_attn.norm.weight",
+      "model.layers.26.input_layernorm.weight",
+      "model.layers.26.post_attention_layernorm.weight",
+      "model.layers.27.self_attn.q_norm.weight",
+      "model.layers.27.self_attn.k_norm.weight",
+      "model.layers.27.input_layernorm.weight",
+      "model.layers.27.post_attention_layernorm.weight",
+      "model.layers.28.linear_attn.norm.weight",
+      "model.layers.28.input_layernorm.weight",
+      "model.layers.28.post_attention_layernorm.weight",
+      "model.layers.29.linear_attn.norm.weight",
+      "model.layers.29.input_layernorm.weight",
+      "model.layers.29.post_attention_layernorm.weight",
+      "model.layers.30.linear_attn.norm.weight",
+      "model.layers.30.input_layernorm.weight",
+      "model.layers.30.post_attention_layernorm.weight",
+      "model.layers.31.self_attn.q_norm.weight",
+      "model.layers.31.self_attn.k_norm.weight",
+      "model.layers.31.input_layernorm.weight",
+      "model.layers.31.post_attention_layernorm.weight",
+      "model.layers.32.linear_attn.norm.weight",
+      "model.layers.32.input_layernorm.weight",
+      "model.layers.32.post_attention_layernorm.weight",
+      "model.layers.33.linear_attn.norm.weight",
+      "model.layers.33.input_layernorm.weight",
+      "model.layers.33.post_attention_layernorm.weight",
+      "model.layers.34.linear_attn.norm.weight",
+      "model.layers.34.input_layernorm.weight",
+      "model.layers.34.post_attention_layernorm.weight",
+      "model.layers.35.self_attn.q_norm.weight",
+      "model.layers.35.self_attn.k_norm.weight",
+      "model.layers.35.input_layernorm.weight",
+      "model.layers.35.post_attention_layernorm.weight",
+      "model.layers.36.linear_attn.norm.weight",
+      "model.layers.36.input_layernorm.weight",
+      "model.layers.36.post_attention_layernorm.weight",
+      "model.layers.37.linear_attn.norm.weight",
+      "model.layers.37.input_layernorm.weight",
+      "model.layers.37.post_attention_layernorm.weight",
+      "model.layers.38.linear_attn.norm.weight",
+      "model.layers.38.input_layernorm.weight",
+      "model.layers.38.post_attention_layernorm.weight",
+      "model.layers.39.self_attn.q_norm.weight",
+      "model.layers.39.self_attn.k_norm.weight",
+      "model.layers.39.input_layernorm.weight",
+      "model.layers.39.post_attention_layernorm.weight",
+      "model.layers.40.linear_attn.norm.weight",
+      "model.layers.40.input_layernorm.weight",
+      "model.layers.40.post_attention_layernorm.weight",
+      "model.layers.41.linear_attn.norm.weight",
+      "model.layers.41.input_layernorm.weight",
+      "model.layers.41.post_attention_layernorm.weight",
+      "model.layers.42.linear_attn.norm.weight",
+      "model.layers.42.input_layernorm.weight",
+      "model.layers.42.post_attention_layernorm.weight",
+      "model.layers.43.self_attn.q_norm.weight",
+      "model.layers.43.self_attn.k_norm.weight",
+      "model.layers.43.input_layernorm.weight",
+      "model.layers.43.post_attention_layernorm.weight",
+      "model.layers.44.linear_attn.norm.weight",
+      "model.layers.44.input_layernorm.weight",
+      "model.layers.44.post_attention_layernorm.weight",
+      "model.layers.45.linear_attn.norm.weight",
+      "model.layers.45.input_layernorm.weight",
+      "model.layers.45.post_attention_layernorm.weight",
+      "model.layers.46.linear_attn.norm.weight",
+      "model.layers.46.input_layernorm.weight",
+      "model.layers.46.post_attention_layernorm.weight",
+      "model.layers.47.self_attn.q_norm.weight",
+      "model.layers.47.self_attn.k_norm.weight",
+      "model.layers.47.input_layernorm.weight",
+      "model.layers.47.post_attention_layernorm.weight",
+      "model.layers.48.linear_attn.norm.weight",
+      "model.layers.48.input_layernorm.weight",
+      "model.layers.48.post_attention_layernorm.weight",
+      "model.layers.49.linear_attn.norm.weight",
+      "model.layers.49.input_layernorm.weight",
+      "model.layers.49.post_attention_layernorm.weight",
+      "model.layers.50.linear_attn.norm.weight",
+      "model.layers.50.input_layernorm.weight",
+      "model.layers.50.post_attention_layernorm.weight",
+      "model.layers.51.self_attn.q_norm.weight",
+      "model.layers.51.self_attn.k_norm.weight",
+      "model.layers.51.input_layernorm.weight",
+      "model.layers.51.post_attention_layernorm.weight",
+      "model.layers.52.linear_attn.norm.weight",
+      "model.layers.52.input_layernorm.weight",
+      "model.layers.52.post_attention_layernorm.weight",
+      "model.layers.53.linear_attn.norm.weight",
+      "model.layers.53.input_layernorm.weight",
+      "model.layers.53.post_attention_layernorm.weight",
+      "model.layers.54.linear_attn.norm.weight",
+      "model.layers.54.input_layernorm.weight",
+      "model.layers.54.post_attention_layernorm.weight",
+      "model.layers.55.self_attn.q_norm.weight",
+      "model.layers.55.self_attn.k_norm.weight",
+      "model.layers.55.input_layernorm.weight",
+      "model.layers.55.post_attention_layernorm.weight",
+      "model.layers.56.linear_attn.norm.weight",
+      "model.layers.56.input_layernorm.weight",
+      "model.layers.56.post_attention_layernorm.weight",
+      "model.layers.57.linear_attn.norm.weight",
+      "model.layers.57.input_layernorm.weight",
+      "model.layers.57.post_attention_layernorm.weight",
+      "model.layers.58.linear_attn.norm.weight",
+      "model.layers.58.input_layernorm.weight",
+      "model.layers.58.post_attention_layernorm.weight",
+      "model.layers.59.self_attn.q_norm.weight",
+      "model.layers.59.self_attn.k_norm.weight",
+      "model.layers.59.input_layernorm.weight",
+      "model.layers.59.post_attention_layernorm.weight",
+      "model.layers.60.linear_attn.norm.weight",
+      "model.layers.60.input_layernorm.weight",
+      "model.layers.60.post_attention_layernorm.weight",
+      "model.layers.61.linear_attn.norm.weight",
+      "model.layers.61.input_layernorm.weight",
+      "model.layers.61.post_attention_layernorm.weight",
+      "model.layers.62.linear_attn.norm.weight",
+      "model.layers.62.input_layernorm.weight",
+      "model.layers.62.post_attention_layernorm.weight",
+      "model.layers.63.self_attn.q_norm.weight",
+      "model.layers.63.self_attn.k_norm.weight",
+      "model.layers.63.input_layernorm.weight",
+      "model.layers.63.post_attention_layernorm.weight",
+      "model.norm.weight"
+    ]
+  },
+  "packages": {
+    "accelerate": "1.14.0",
+    "bitsandbytes": "0.50.1",
+    "causal-conv1d": "1.7.0",
+    "fla-core": "0.5.2",
+    "flash-linear-attention": "0.5.2",
+    "huggingface-hub": "1.28.0",
+    "liger-kernel": "0.8.2",
+    "peft": "0.20.0",
+    "torch": "2.8.0+cu128",
+    "transformers": "5.15.0"
+  },
+  "resume": false,
+  "run_id": "pilot-live-v15-exhaustive-unit-r16-lr2e5-seed38-v1",
+  "save_steps": 100,
+  "seed": 38,
+  "smoke_longest": false,
+  "tool_name": "submit_facts",
+  "tools_sha256": "e02cf3c20831eed00a05598f42f324f9112f7c283b9cd0cef43af96e294c93ab",
+  "tools_transport_sha256": "8945d9bfd5e68ea50537c4cae5476fcfa9efd4e99ab80c9b71bc4bfcbd5826b1",
+  "train_files": [
+    {
+      "name": "train.jsonl",
+      "sha256": "1b3630d00d11476c7f1e941a60a442a5eb5eb57dd882b7d188ef9b85d461e738"
+    }
+  ],
+  "train_sha256": "1b3630d00d11476c7f1e941a60a442a5eb5eb57dd882b7d188ef9b85d461e738",
+  "train_stats": {
+    "components": [
+      {
+        "dropped": 0,
+        "kept": 4255,
+        "loss_mode_counts": {
+          "full_assistant": 4255
+        },
+        "max_loss_tokens": 639,
+        "max_tokens": 3109,
+        "mean_loss_tokens": 147.76,
+        "mean_tokens": 1363.21,
+        "min_loss_tokens": 44,
+        "min_tokens": 964,
+        "path": "/workspace/donto-qwen80/data/pilot-training-live-v15-exhaustive-unit-v3/train.jsonl",
+        "rows": 4255
+      }
+    ],
+    "dropped": 0,
+    "kept": 4255,
+    "loss_mode_counts": {
+      "full_assistant": 4255
+    },
+    "max_loss_tokens": 639,
+    "max_tokens": 3109,
+    "mean_loss_tokens": 147.76,
+    "mean_tokens": 1363.21,
+    "min_loss_tokens": 44,
+    "min_tokens": 964,
+    "rows": 4255
+  },
+  "trainer_source_sha256": "994d3f888fe3d18f115ac050d59aa68a524ef160dc7f2073a83d4a8568b82d79",
+  "validation_sha256": "c360e8296e5ec7973f2fab8416d0b30e2aa850fe6b2cc618bb3fa52894d818b2",
+  "validation_stats": {
+    "dropped": 0,
+    "kept": 1085,
+    "loss_mode_counts": {
+      "full_assistant": 1085
+    },
+    "max_loss_tokens": 627,
+    "max_tokens": 2198,
+    "mean_loss_tokens": 140.61,
+    "mean_tokens": 1346.18,
+    "min_loss_tokens": 44,
+    "min_tokens": 964,
+    "rows": 1085
+  },
+  "warmup_ratio": 0.05
+}

chat_template.jinja

@@ -0,0 +1,170 @@
+{%- set image_count = namespace(value=0) %}
+{%- set video_count = namespace(value=0) %}
+{%- macro render_content(content, do_vision_count, is_system_content=false) %}
+    {%- if content is string %}
+        {{- content }}
+    {%- elif content is iterable and content is not mapping %}
+        {%- for item in content %}
+            {%- if 'image' in item or 'image_url' in item or item.type == 'image' %}
+                {%- if is_system_content %}
+                    {{- raise_exception('System message cannot contain images.') }}
+                {%- endif %}
+                {%- if do_vision_count %}
+                    {%- set image_count.value = image_count.value + 1 %}
+                {%- endif %}
+                {%- if add_vision_id %}
+                    {{- 'Picture ' ~ image_count.value ~ ': ' }}
+                {%- endif %}
+                {{- '<|vision_start|><|image_pad|><|vision_end|>' }}
+            {%- elif 'video' in item or item.type == 'video' %}
+                {%- if is_system_content %}
+                    {{- raise_exception('System message cannot contain videos.') }}
+                {%- endif %}
+                {%- if do_vision_count %}
+                    {%- set video_count.value = video_count.value + 1 %}
+                {%- endif %}
+                {%- if add_vision_id %}
+                    {{- 'Video ' ~ video_count.value ~ ': ' }}
+                {%- endif %}
+                {{- '<|vision_start|><|video_pad|><|vision_end|>' }}
+            {%- elif 'text' in item %}
+                {{- item.text }}
+            {%- else %}
+                {{- raise_exception('Unexpected item type in content.') }}
+            {%- endif %}
+        {%- endfor %}
+    {%- elif content is none or content is undefined %}
+        {{- '' }}
+    {%- else %}
+        {{- raise_exception('Unexpected content type.') }}
+    {%- endif %}
+{%- endmacro %}
+{%- if not messages %}
+    {{- raise_exception('No messages provided.') }}
+{%- endif %}
+{%- set reasoning_instructions = '' %}
+{%- if enable_thinking is undefined or enable_thinking is true %}
+    {%- set resolved_reasoning_effort = reasoning_effort|default('xhigh') %}
+    {%- if resolved_reasoning_effort not in ('xhigh', 'medium', 'low') %}
+        {{- raise_exception('Unexpected reasoning effort ' ~ reasoning_effort ~ '. Supported types are xhigh (default), medium, and low.') }}
+    {%- endif %}
+    {%- if resolved_reasoning_effort == 'xhigh' %}
+        {%- set reasoning_instructions = 'Reasoning effort is set to xhigh. Please think carefully through the task, validate key assumptions, consider plausible alternatives, and prioritize correctness, consistency, and clarity in the final answer.' %}
+    {%- elif resolved_reasoning_effort == 'low' %}
+        {%- set reasoning_instructions = 'Reasoning effort is set to low. Keep your thinking brief and focused, moving directly to the conclusion without unnecessary elaboration.' %}
+    {%- endif %}
+{%- endif %}
+{%- if tools and tools is iterable and tools is not mapping %}
+    {{- '<|im_start|>system\n' }}
+    {%- if reasoning_instructions %}
+        {{- reasoning_instructions + '\n\n' }}
+    {%- endif %}
+    {{- "# Tools\n\nYou have access to the following functions:\n\n<tools>" }}
+    {%- for tool in tools %}
+        {{- "\n" }}
+        {{- tool | tojson }}
+    {%- endfor %}
+    {{- "\n</tools>" }}
+    {{- '\n\nIf you choose to call a function ONLY reply in the following format with NO suffix:\n\n<tool_call>\n<function=example_function_name>\n<parameter=example_parameter_1>\nvalue_1\n</parameter>\n<parameter=example_parameter_2>\nThis is the value for the second parameter\nthat can span\nmultiple lines\n</parameter>\n</function>\n</tool_call>\n\n<IMPORTANT>\nReminder:\n- Function calls MUST follow the specified format: an inner <function=...></function> block must be nested within <tool_call></tool_call> XML tags\n- Required parameters MUST be specified\n- You may provide optional reasoning for your function call in natural language BEFORE the function call, but NOT after\n- If there is no function call available, answer the question like normal with your current knowledge and do not tell the user about function calls\n</IMPORTANT>' }}
+    {%- if messages[0].role == 'system' %}
+        {%- set content = render_content(messages[0].content, false, true)|trim %}
+        {%- if content %}
+            {{- '\n\n' + content }}
+        {%- endif %}
+    {%- endif %}
+    {{- '<|im_end|>\n' }}
+{%- else %}
+    {%- if messages[0].role == 'system' %}
+        {%- set content = render_content(messages[0].content, false, true)|trim %}
+        {%- if content %}
+            {{- '<|im_start|>system\n' + (reasoning_instructions + '\n\n' if reasoning_instructions else '')  + content + '<|im_end|>\n' }}
+        {%- elif reasoning_instructions %}
+            {{- '<|im_start|>system\n' + reasoning_instructions + '<|im_end|>\n' }}
+        {%- endif %}
+    {%- elif reasoning_instructions %}
+        {{- '<|im_start|>system\n' + reasoning_instructions + '<|im_end|>\n' }}
+    {%- endif %}
+{%- endif %}
+{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}
+{%- for message in messages[::-1] %}
+    {%- set index = (messages|length - 1) - loop.index0 %}
+    {%- if ns.multi_step_tool and message.role == "user" %}
+        {%- set content = render_content(message.content, false)|trim %}
+        {%- if not(content.startswith('<tool_response>') and content.endswith('</tool_response>')) %}
+            {%- set ns.multi_step_tool = false %}
+            {%- set ns.last_query_index = index %}
+        {%- endif %}
+    {%- endif %}
+{%- endfor %}
+{%- if ns.multi_step_tool %}
+    {{- raise_exception('No user query found in messages.') }}
+{%- endif %}
+{%- for message in messages %}
+    {%- set content = render_content(message.content, true)|trim %}
+    {%- if message.role == "system" %}
+        {%- if not loop.first %}
+            {{- raise_exception('System message must be at the beginning.') }}
+        {%- endif %}
+    {%- elif message.role == "user" %}
+        {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {%- set reasoning_content = '' %}
+        {%- if message.reasoning_content is string %}
+            {%- set reasoning_content = message.reasoning_content %}
+        {%- endif %}
+        {%- set reasoning_content = reasoning_content|trim %}
+        {%- if preserve_thinking is undefined or preserve_thinking is true or loop.index0 > ns.last_query_index %}
+            {{- '<|im_start|>' + message.role + '\n<think>\n' + reasoning_content + '\n</think>\n\n' + content }}
+        {%- else %}
+            {{- '<|im_start|>' + message.role + '\n' + content }}
+        {%- endif %}
+        {%- if message.tool_calls and message.tool_calls is iterable and message.tool_calls is not mapping %}
+            {%- for tool_call in message.tool_calls %}
+                {%- if tool_call.function is defined %}
+                    {%- set tool_call = tool_call.function %}
+                {%- endif %}
+                {%- if loop.first %}
+                    {%- if content|trim %}
+                        {{- '\n\n<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                    {%- else %}
+                        {{- '<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                    {%- endif %}
+                {%- else %}
+                    {{- '\n<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                {%- endif %}
+                {%- if tool_call.arguments is defined and tool_call.arguments != '' %}
+                    {%- for args_name, args_value in tool_call.arguments|items %}
+                        {{- '<parameter=' + args_name + '>\n' }}
+                        {%- set args_value = args_value | string if args_value is string else args_value | tojson | safe %}
+                        {{- args_value }}
+                        {{- '\n</parameter>\n' }}
+                    {%- endfor %}
+                {%- endif %}
+                {{- '</function>\n</tool_call>' }}
+            {%- endfor %}
+        {%- endif %}
+        {{- '<|im_end|>\n' }}
+    {%- elif message.role == "tool" %}
+        {%- if loop.previtem and loop.previtem.role != "tool" %}
+            {{- '<|im_start|>user' }}
+        {%- endif %}
+        {{- '\n<tool_response>\n' }}
+        {{- content }}
+        {{- '\n</tool_response>' }}
+        {%- if not loop.last and loop.nextitem.role != "tool" %}
+            {{- '<|im_end|>\n' }}
+        {%- elif loop.last %}
+            {{- '<|im_end|>\n' }}
+        {%- endif %}
+    {%- else %}
+        {{- raise_exception('Unexpected message role.') }}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|im_start|>assistant\n' }}
+    {%- if enable_thinking is defined and enable_thinking is false %}
+        {{- '<think>\n\n</think>\n\n' }}
+    {%- else %}
+        {{- '<think>\n' }}
+    {%- endif %}
+{%- endif %}

docs/FREEZE-RECEIPT-20260825.md

@@ -0,0 +1,38 @@
+# Synthetic Gold 100 Graph-First V1 — Freeze Receipt
+
+Frozen at `2026-08-25T10:57:55Z`, before any V15 inference over this suite.
+
+## Scope
+
+- 100 unique graph-first synthetic documents
+- 2,718 structured extraction decisions
+- 5,338 row-level target facts
+- 5,288 document-level unique `(s,p,o)` facts
+- 45–59 unique gold facts per document; mean 52.88
+- 73 blind parent templates
+- domains and constructions: aviation incidents, diplomatic quotations, mining and clinical-scheduling conditionals, argument chains, nested beliefs, procurement, appointments, and coalitions
+
+## Leakage audit against V15
+
+Compared with `pilot-training-live-v15-exhaustive-unit-v3`:
+
+- blind unique source hashes: 100
+- V15 train/validation unique source hashes: 240
+- source-hash overlap: 0
+- blind unique parent templates: 73
+- V15 train/validation unique parent templates: 102
+- parent-template overlap: 0
+- four operator/admission fixture hashes: hard excluded
+
+The generator's structural audit passed all checks in `audit.json`. The blind suite was not opened during construction or audit.
+
+## Frozen hashes
+
+- `sealed-blind.jsonl`: `6f16c87d06c2a0b850321b8750e54a63364937cfa2823f34f698f6c00f160f1c`
+- `validator-sealed-blind.jsonl`: `19a293e9d3bb142cb737b9fbef311508b7867934bd2c00c80adb13ba10342a5e`
+- `tools.json`: `e02cf3c20831eed00a05598f42f324f9112f7c283b9cd0cef43af96e294c93ab`
+- `validator-tools.json`: `2c85d9fe1da85b9b6a00698ebd34776adfa453d12f3eef5397ab8149c5081549`
+- `manifest.json`: `dd46f14c743880a1a4692406ae6fa88db80dc06e89ff70ae85b8d7fb00b8bcd5`
+- `audit.json`: `4b04ecc9fa3e3ccddabe14189396cf38cf555c17088565e57e8f2dd40835a3e8`
+
+The canonical extraction score is exact structured fact recall/precision over all 2,718 decisions, reported globally, by task, template, and document. This is intentionally stricter than a one-prompt-per-document count test.

docs/LAUNCH-RECEIPT.md

@@ -0,0 +1,25 @@
+# V15 full-pilot launch receipt
+
+- UTC launch: 2026-08-25T00:34Z
+- Run ID: `pilot-live-v15-exhaustive-unit-r16-lr2e5-seed38-v1`
+- RunPod pod: `wydkrrvhi4iqbv`
+- GPU: one NVIDIA GeForce RTX 3090, 24,576 MiB
+- Remote trainer PID: `14504`
+- Base: `/workspace/models/Qwen3.8-27B-BF16`, pinned revision `1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0`
+- Corpus: `/workspace/donto-qwen80/data/pilot-training-live-v15-exhaustive-unit-v3`
+- Trainer SHA-256: `994d3f888fe3d18f115ac050d59aa68a524ef160dc7f2073a83d4a8568b82d79`
+- Launcher SHA-256: `236bac7e50741e934fc42e4f6aa512f32b0b14985e908bb60520573c9a81cfee`
+- Recipe: rank 16, no rsLoRA, LR 2e-5, seed 38, gradient accumulation 8, one epoch, warmup ratio 0.05, maximum length 4096, checkpoint every 100 steps.
+- Expected optimizer steps: 532 from 4,255 training rows.
+- Blind/admission sources: not opened.
+
+This receipt records launch only. It is not evidence that training completed or that the resulting model passed quality admission.
+
+## Verified live state
+
+- First optimizer step completed with loss 0.1898; subsequent steady-state steps averaged about 36–38 seconds.
+- At step 29/532 the process remained alive, GPU utilization was 100%, VRAM was 21,576 MiB, temperature was 63°C, and the most recent loss was 0.1105.
+- `donto-qwen-v15-checkpoint-mirror.service` is active and will checksum-mirror every completed 100-step checkpoint.
+- `donto-qwen-v15-final-artifact-mirror.service` is active and will copy final artifacts only after the trainer exits with `TRAINING-COMPLETE.json` and a final adapter.
+- The complete physical longest-row smoke bundle is already checksum-mirrored at the sibling `smoke-pilot-live-v15-exhaustive-unit-r16-longest-v1/` directory.
+- Production extraction remains paused; sealed admission fixtures remain unopened.

docs/PILOT-LIVE-V11-REPAIR-REJECTION-20260823.md

@@ -0,0 +1,55 @@
+# Pilot live V11 repair rejection
+
+The candidate `pilot-live-v11-repair-r16-lr2e5-seed38-v1` is rejected. It is
+retained as a negative result and must not be used as the weight base for a
+later adapter.
+
+## Source-disjoint structured validation
+
+- artifact: `results/pilot-live-v11-repair-r16-v1-validation-full428.json`
+- SHA-256: `2bf90e6e09eb5c5b75c872c463f5715cf624afe572e530d992a3ab64f2300cb1`
+- 428/428 successful structured tool calls
+- 1,584/1,792 exact target facts recovered (88.39% recall)
+- 1,584/1,780 exact predicted facts matched (88.99% precision)
+- zero malformed, duplicate-key, or reflexive items
+- bootstrap 95% lower recall bound: 77.10%
+
+Recall exceeded the E2 floor, but exact precision missed the declared 90% gate.
+The largest mismatch cluster was stable entity/event identifier divergence; this
+does not excuse the failed predeclared gate.
+
+## Real `llm_nerd` development diagnosis
+
+- artifact: `results/pilot-live-v11-repair-r16-v1-llm-nerd-expansion4-no-validator-repeat1.json`
+- SHA-256: `c2a6f7578a3967b95182cb0e4905f46ad35fa925a05ba25b8c6829dd8340e77c`
+- finite completion: yes
+- facts: 54
+- stage errors: zero
+- conservative semantic score: 16/46 substantive, including 12/27 direct and
+  4/19 derived; 12/36 structural; 28/82 whole gold
+- seven clear role, qualifier, modality, or layer errors
+- many structural alias duplicates
+
+The hash-bound adjudication is
+`adjudication/pilot-live-v11-repair-llm-nerd-ledger-v1.json`.
+
+## Corpus-integrity finding
+
+The V11 `abundance-argument` generator preserved no literal fixture text, but it
+reproduced the development fixture's complete semantic narrative arc too closely:
+belief and scope, unreliable dealings, vendor and lender unavailability,
+leadership selection, threat countering, alliance dependency, attacks and
+diminishing, and a self-defeating conclusion. That violates the plan's explicit
+ban on close paraphrases. The fact that the candidate still failed does not make
+the contamination acceptable.
+
+The template is quarantined as non-compiling negative-result code. The V12
+repair replaces it with four independent capability families whose stories do
+not combine the fixture arc: belief/evidence, procurement consequences,
+appointment/quotation, and coalition strategy.
+
+## Next action
+
+Train a fresh-stock V12 pilot on the clean, source-disjoint corpus. Do not train
+the validator yet: the principal failure remains missing semantic families, and
+a validator cannot add them.

docs/PILOT-LIVE-V12-CLEAN-REJECTION-20260823.md

@@ -0,0 +1,42 @@
+# Pilot live-v12 clean rejection
+
+`pilot-live-v12-clean-d1weighted-r16-lr2e5-seed38-v1` is rejected and must not
+be promoted or used as the starting weights for another candidate.
+
+The serving-path corpus validation completed all 346 rows with no failed tool
+calls. It achieved 900/1260 exact recall (71.43%), above the pilot advance floor
+of 65%, with 81.12% exact structural recall. The 30-row direct semantic review
+found 97/99 supported predictions (97.98%). Those results establish that the
+adapter is structurally usable and that the clean curriculum is learnable.
+
+The real `llm_nerd` development target nevertheless failed decisively. Three
+bounded deterministic repeats emitted 47, 46, and 47 facts. All were finite and
+had zero stage errors, but no 47-fact trajectory can meet the 67/82 whole-gold
+gate. Conservative one-to-one review of repeat 1 scored:
+
+- substantive: 18/46;
+- direct: 13/27;
+- derived: 5/19;
+- structural: 7/36;
+- whole gold: 25/82;
+- unique-supported precision: 32/44 (72.73%);
+- clear role, scope, reflexive, or support errors: 12.
+
+The failure is not an open-ended stopping loop. Every unit and expansion layer
+finished. It is a coverage-granularity and role-control failure: every D0 unit
+returned exactly two facts; structural expansion stopped after ten facts; and
+global D1 generation introduced backwards or semantic-self relations.
+
+The next experiment will retain the finite controller but make coverage state
+local and explicit: bounded per-unit D0 follow-ups, source-bound D1 expansion,
+and structural batches over the dynamically discovered entity inventory. If
+that architectural ceiling remains below the gold gate, the resulting
+source-bound misses become clean on-policy supervision for a fresh-stock V13
+coverage adapter. The sealed admission documents and blind suite remain unopened.
+
+Evidence:
+
+- full validation: `results/pilot-live-v12-clean-d1weighted-r16-v1-validation-full346.json`;
+- first-30 review: `adjudication/pilot-live-v12-clean-validation-first30-semantic-review-v1.json`;
+- real-document ledger: `adjudication/pilot-live-v12-clean-llm-nerd-ledger-v1.json`;
+- complete canary log: `runpod-runs/pilot-live-v12-clean-d1weighted-r16-lr2e5-seed38-v1/serving-evidence/pilot-live-v12-clean-d1weighted-r16-v1-bnb-canary.log`.

docs/PILOT-LIVE-V13-FULL746-SEMANTIC-MISMATCH-AUDIT-20260824.md

@@ -0,0 +1,82 @@
+# V13 full-746 semantic mismatch audit
+
+**Candidate:** `pilot-live-v13-controller-aligned-r16-v1`  
+**Evaluator:** `implementation-snapshot-live-v23/donto-qwen38-eval-corpus`  
+**Result:** `pilot-live-v13-controller-aligned-r16-v1-validation-full746.json`  
+**Scope:** non-blind source-disjoint pilot validation only
+
+## Measured result
+
+The run completed 746/746 forced `submit_facts` calls with no failed or malformed
+call, no invalid fact object, no duplicate exact item, and no reflexive entity
+item.
+
+```text
+target facts                  1260
+predicted facts               1252
+exact one-to-one matches      1128
+exact recall                  0.8952380952380953
+exact precision               0.9009584664536742
+bootstrap 95% recall floor    0.8696338577824516
+mean document exact recall    0.8998281895665682
+tool-call success             1.0
+```
+
+The 1128 exact matches are themselves source-bound reviewed targets. Because all
+124 non-exact predictions are already treated as unsupported in the exact
+denominator, `1128 / 1252 = 0.900958` is a conservative supported-precision
+lower bound. Giving semantic credit to a supported free-predicate synonym or
+stable entity alias can only raise that value. The pilot therefore clears the
+declared 0.90 precision floor without relying on generous semantic matching.
+
+## Complete non-exact-row review
+
+All 86 rows containing a non-exact target or prediction were read directly
+against their target and source-bound task prompt. Classification is at the row
+level, so a mixed row may contain both a useful supported alternative and an
+omission or incorrect role.
+
+| Classification | Count | Row indexes |
+|---|---:|---|
+| Supported alternate, stable alias, or harmless structural granularity | 40 | 63, 95, 139, 152, 158, 183, 190, 245, 248, 251, 280, 284, 293, 347, 354, 366, 369, 377, 386, 439, 465, 469, 504, 505, 512, 523, 525, 531, 547, 556, 587, 590, 657, 669, 690, 714, 716, 720, 727, 742 |
+| Omission only; no extra prediction to assess | 10 | 130, 205, 320, 367, 393, 467, 693, 705, 712, 740 |
+| Clear redundant, unsupported, role-direction, or scope error | 16 | 5, 126, 171, 176, 317, 332, 334, 336, 436, 520, 544, 621, 707, 709, 730, 736 |
+| Mixed supported alternative plus omission or questionable relation | 20 | 96, 135, 150, 167, 169, 179, 232, 281, 311, 328, 339, 363, 420, 468, 503, 542, 552, 653, 701, 739 |
+
+These four sets contain all 86 mismatch rows exactly once.
+
+## Failure patterns that remain real
+
+- Row 5 drifts from Oona's requested recording action to an unrelated automatic
+  behavior.
+- Rows 171, 176, 332, 336, 436, 520, 544, 621, 709, and 736 contain a reversed
+  or substituted subject/object role.
+- Rows 126, 317, and 334 emit a structurally invalid label or a redundant/newness
+  failure.
+- Rows 707 and 730 invent the wrong causal or strategic direction.
+- D2 is the weakest exact stage: 73/100 target facts, with 0.8022 exact
+  precision. Its misses concentrate in qualification, discourse scope, and
+  reified strategic conclusions rather than tool formatting.
+
+## Stage breakdown
+
+| Task | Recall | Precision | Calls |
+|---|---:|---:|---:|
+| D0 | 0.8843 | 0.8843 | 108 |
+| D1 unit | 0.8973 | 0.8912 | 190 |
+| D2 | 0.7300 | 0.8022 | 50 |
+| Metadata | 0.9643 | 0.9878 | 24 |
+| Residual | 0.8333 | 0.8219 | 48 |
+| Residual unit | 0.9060 | 0.8993 | 230 |
+| Structural batch | 0.9335 | 0.9360 | 72 |
+
+## Gate decision
+
+V13 passes the controller-aligned pilot structured gate and may proceed to the
+three deterministic non-blind `llm_nerd` development trajectories. This is not
+acceptance: the real-document gate independently requires 38/46 substantive,
+22/27 direct, 16/19 derived, 29/36 structural, 67/82 whole-gold, at least 0.90
+unique supported precision, no more than one clear role/reflexive error, and
+finite valid completion in every repeat.
+
+The blind holdout and three synthetic admission documents remain unopened.

docs/PILOT-LIVE-V14-DENSE-UNIT-REJECTION-20260825.md

@@ -0,0 +1,30 @@
+# Pilot live-v14 dense-unit rejection
+
+`pilot-live-v14-dense-unit-r16-lr2e5-seed38-v1` and its checkpoint/scaling variants are rejected. They must not be promoted or used as starting weights for another candidate.
+
+V14 learned its supervised contract extremely well. The complete source-disjoint validation run finished all 1,214 rows with no failed tool calls, 2,071/2,092 exact recall (98.996%), 2,071/2,085 exact precision (99.329%), and a 98.412% bootstrap lower bound. Checkpoint 500 scored 46/46 on its first 30 validation rows. These results establish learnability, formatting reliability, and a useful stopping point near 500 steps.
+
+The real `llm_nerd` development document still failed decisively. The best breadth probe was checkpoint 500 under the frozen live-v26 controller. It completed finitely with no stage errors and emitted 75 facts, but conservative one-to-one review scored:
+
+- substantive: 15/46;
+- direct: 10/27;
+- derived: 5/19;
+- structural: 17/36;
+- whole gold: 32/82;
+- unique-supported precision: 47/65 (72.31%);
+- clear role, scope, reflexive, or support errors: 18.
+
+The 0.75 LoRA-scale ablation also failed: its first 30 supervised rows fell to 41/46 exact recall and 41/47 precision, while its real-document run emitted only 61 facts and retained serious role errors. Scale 0.50 collapsed further to 22/46 recall and 22/73 precision. LoRA scaling is therefore not the remedy.
+
+The failure is a curriculum-transfer mismatch. V14 capped dense D0 targets at four facts, taught empty local residuals, and rewarded clean short pages. The live controller asks for more exhaustive atomic coverage. Checkpoint 500 increased raw volume, but the added facts were mainly aliases, duplicates, malformed inverses, and incorrect role assignments—not the missing vendor, bank, causal, alliance-participant, behavior-mode, discourse, and structural families.
+
+V15 consequently starts fresh from the pinned stock base. It teaches three-to-seven-fact dense D0 pages, leaves one supported dense-unit fact for a nonempty residual recovery pass, follows that with an explicit empty terminal, increases local D1 capacity, and broadens source/template families. No V14 weights are reused. The three synthetic admission documents and sealed blind suite remain unopened.
+
+Evidence:
+
+- full supervised validation: `results/pilot-live-v14-dense-unit-r16-v1-validation-full1214.json`;
+- checkpoint-500 supervised probe: `results/pilot-live-v14-dense-unit-r16-step500-validation-first30.json`;
+- checkpoint-500 real trajectory: `results/pilot-live-v14-dense-unit-r16-step500-controller-livev26-llm-nerd-repeat1.json`;
+- manual ledger: `adjudication/pilot-live-v14-dense-unit-step500-llm-nerd-ledger-v1.json`;
+- 0.75-scale validation and trajectory: `results/pilot-live-v14-dense-unit-r16-scale075-validation-first30.json` and `results/pilot-live-v14-dense-unit-r16-scale075-controller-livev26-llm-nerd-repeat1.json`;
+- 0.50-scale validation: `results/pilot-live-v14-dense-unit-r16-scale050-validation-first30.json`.

docs/PILOT-LIVE-V15-EXHAUSTIVE-UNIT-CORPUS-AUDIT-20260825.md

@@ -0,0 +1,51 @@
+# Pilot live-v15 exhaustive-unit corpus audit
+
+The canonical V15 pilot corpus is admitted for development training. This is a corpus/runtime-contract decision only; it is not model-quality admission.
+
+## Immutable inputs
+
+- source corpus: `pilot-corpus-live-v15-exhaustive-unit-v3/`;
+- trainer corpus: `pilot-training-live-v15-exhaustive-unit-v3/`;
+- train: 4,255 rows, 9,013 target facts, 200 source families;
+- validation: 1,085 rows, 2,140 target facts, 40 source-disjoint families;
+- source-family overlap: 0;
+- parent-template overlap: 0;
+- blind rows used: 0;
+- excluded operator/admission source hashes: 4.
+
+The training split combines 4,123 synthetic rows from 152 synthetic documents with 132 reviewed-real rows from 48 source families. Validation contains 40 different synthetic source/template families. The `llm_nerd` source, Trump/China/Russia development literals, three synthetic admission documents, and sealed blind suite are absent from the trainer inputs.
+
+## V15 task-shape correction
+
+The prior V14 curriculum taught clean but terse pages. V15 changes the supervised state distribution:
+
+- dense D0 pages contain three to seven direct facts rather than being capped at four;
+- every dense unit deliberately leaves one supported fact for a nonempty local residual recovery pass;
+- the next identical residual audit is explicitly empty, teaching finite exhaustion after recovery;
+- local D1 pass one may contain up to twelve licensed views, followed by an empty terminal pass;
+- half of synthetic documents come from the new incident, quotation, conditional, argument, belief, procurement, appointment, and coalition depth families;
+- reviewed-real rows remain in the source-balanced mixture.
+
+Training-task counts for the canonical train split include 646 D0 rows with 2,641 facts, 893 local residual rows with 247 recovery facts and 646 explicit empty terminals, 1,159 local D1 rows with 988 facts and 646 empty terminals, plus document metadata, global residual, D2, structural, and reviewed-real rows.
+
+## Hashes
+
+- `train.jsonl`: `1b3630d00d11476c7f1e941a60a442a5eb5b57dd882b7d188ef9b85d461e738`;
+- `validation.jsonl`: `c360e8296e5ec7973f2fab8416d0b30e2aa850fe6b2cc618bb3fa52894d818b2`;
+- `tools.json`: `e02cf3c20831eed00a05598f42f324f9112f7c283b9cd0cef43af96e294c93ab`;
+- derived `manifest.json`: `26e3376dc6b99f8b1392a868dd0365a65ae0f1e4e20721cf994ad2f46d1a3231`;
+- derived `audit.json`: `7cd156217b16636a2c918ddd737dd33b0ee2a712e773dab0984a0dcd91d52c7d`;
+- rendered-row audit: `03d5c82f6ea3deb387e80a1100d9715af742363d5579a50959d1ede560cb61b3`;
+- exact trainer run contract: `0a6e1976e11147a198598aab158523f0fc676adde260a496b167e67dbc4e8860`;
+- live tokenizer contract receipt: `2eaef29c9158852e6376facd89bf0f4ecc108dcd42558e36e072f7fbab841144`;
+- trainer source: `994d3f888fe3d18f115ac050d59aa68a524ef160dc7f2073a83d4a8568b82d79`.
+
+## Render and physical checks
+
+All 4,255 train and 1,085 validation rows render below the 4,096-token contract. Train spans 964–3,109 tokens; validation spans 964–2,198; over-limit rows are zero. The live vLLM tokenizer reproduced the first validation prompt at exactly 1,133 tokens with the expected token hash.
+
+The physical longest-row smoke ran on the approved RTX 3090 using the exact pinned training environment. It completed one real optimizer step over 3,109- and 3,020-token training rows, ran validation at 2,198 tokens, produced normal loss/eval metrics, adapted exactly 496 language modules with 116,727,808 trainable parameters and 992 LoRA tensors, touched no vision module, wrote `TRAINING-COMPLETE.json`, and returned GPU allocation to idle. Its complete 1,192,924,351-byte tree is independently checksum-mirrored at `runpod-runs/smoke-pilot-live-v15-exhaustive-unit-r16-longest-v1/`.
+
+## Decision
+
+The V15 corpus and physical runtime contract pass the development-training gate. They do not establish semantic transfer. The full seed-38 pilot must still pass supervised serving checks and the manually adjudicated real-document gold gate before any multi-seed or sealed admission work.

docs/PILOT-LIVE-V15-EXHAUSTIVE-UNIT-POSTTRAIN-RUNBOOK.md

@@ -0,0 +1,43 @@
+# V15 exhaustive-unit pilot post-training runbook
+
+This runbook is bound to `pilot-live-v15-exhaustive-unit-r16-lr2e5-seed38-v1`. It does not enable a production lane or open the three synthetic admission documents or sealed 100-document suite.
+
+## 1. Completion and durability
+
+Require all 532 optimizer steps, epoch validation, explicit final validation, `TRAINING-COMPLETE.json`, clean exit of the exact trainer PID, and GPU allocation returned near zero. Checkpoint 100/200/300/400/500 directories must be mirrored to the mounted drive with stable before/after remote hashes and exact local reconciliation. Copy and reconcile the complete final run tree and detached training log separately.
+
+Treat a PID exit without `TRAINING-COMPLETE.json` as failure. Do not infer success from a final adapter directory alone.
+
+## 2. Tensor-preserving conversion
+
+Convert the trainer-native final adapter with the conditional vLLM remapper. Require 992 tensors, the conditional language-model key root, all twelve expected target suffixes, no vision tensor, and `tensor_values_changed=false`. Perform the same conversion for checkpoint 500 so the planned early-stop comparison remains available.
+
+## 3. Isolated BF16 canary
+
+Start the isolated BF16-bitsandbytes vLLM canary with one adapter at a time, maximum model length 4,096, maximum sequences 4, prefix caching enabled, thinking disabled, greedy decoding, and zero presence/frequency penalties. Verify `/v1/models`, one forced `submit_facts` tool call, bounded memory, and the exact tokenizer contract before corpus evaluation.
+
+Production extraction controls remain paused throughout development evaluation.
+
+## 4. Non-blind evaluation sequence
+
+For the final adapter:
+
+1. Evaluate the first 30 V15 validation rows with concurrency 4 and 1,024 output tokens.
+2. Require 30/30 valid tool calls and no reflexive or malformed outputs. Continue to all 1,085 rows only if exact recall is at least 80% and exact precision is at least 90%.
+3. Evaluate the full V15 validation set and preserve task, stratum, and source-family breakdowns.
+4. Run `llm_nerd` through the frozen live-v27 controller with metadata, residual, D1, D2, and structural stages enabled and validator disabled.
+5. Run three deterministic repeats. Require bounded completion and zero malformed calls in all three.
+6. Manually adjudicate every unique prediction one-to-one against the frozen 46 substantive and 36 structural families.
+
+The real-document gate is at least 38/46 substantive, including 22/27 direct and 16/19 derived, at least 29/36 structural and 67/82 whole, at least 90% unique-supported precision, and no more than one clear role-direction or reflexive error.
+
+If the final adapter misses because the last 32 steps made it terser, run the same sequence on checkpoint 500. This comparison was fixed before seeing V15 outputs and does not authorize arbitrary checkpoint fishing.
+
+## 5. Branch decision
+
+- If both final and checkpoint 500 fail substantive recall, reject V15 and use only non-blind trajectories to correct the coverage curriculum. Do not inherit rejected weights.
+- If recall passes but precision or role direction fails, build the source-bound validator/repair adapter from on-policy mistakes and rerun the same development gate.
+- If one V15 checkpoint passes every development gate, freeze its recipe and train seeds 38, 73, and 211 sequentially from the pinned stock base on the same approved RTX 3090. At least two of three must pass.
+- Only after recipe freeze and multi-seed confirmation may the three synthetic admission documents and sealed suite be opened once.
+
+Quantization and production packaging begin only after BF16 semantic admission. The RTX 3090 quantized system must remain within three recall points of its BF16 result.

docs/QWEN-DONTO-80-PERCENT-GOLD-PLAN-20260822.md

@@ -0,0 +1,638 @@
+# Qwen Donto plan for at least 80% gold coverage
+
+**Date:** 2026-08-22  
+**Decision requested:** how to move the local Donto Qwen3.8-27B system from its
+current low-recall predicate extraction to at least 80% of the adjudicated Codex
+gold.  
+**Scope of this document:** corpus construction, retraining, inference
+orchestration, semantic validation, evaluation, quantization, deployment, and
+cost. This is a plan and research record. No training or production-lane restart
+was performed while producing it.
+
+## Executive verdict
+
+The 80% target is credible, but not by training another stop boolean or merely
+asking the current one-page prompt to be more exhaustive.
+
+The shortest credible route is:
+
+1. expose every source unit to the extractor;
+2. train a new coverage adapter on direct, inverse/decomposed, discourse, and
+   structural tasks that the current corpus does not teach;
+3. train a small second adapter to reject or repair reversed roles and scope
+   errors using candidates generated by Qwen itself;
+4. let a finite Rust controller, not Qwen's `more_supported_facts` prediction,
+   determine when the chunk is complete;
+5. admit the quantized RTX 3090 system only after semantic one-to-one evaluation
+   reaches the declared gate.
+
+The primary gate is **at least 38/46 substantive gold rows** on the `llm_nerd`
+fixture: at least 22/27 directly licensed rows and 16/19 derived or interpretive
+rows. This is 82.6%, giving one row of margin above the literal 37/46 threshold.
+
+The literal whole-gold companion gate is **at least 67/82 rows**: the substantive
+gate plus at least 29/36 structural type/label rows. Structural rows are reported
+separately so mechanically expanding labels cannot disguise poor predicate
+understanding.
+
+This fixture is a development diagnostic, not release evidence. The same recipe
+must also average at least 80% substantive semantic recall on a source-disjoint
+suite while maintaining at least 90% unique supported-fact precision.
+
+## Why 80% is attainable from the evidence already collected
+
+The best no-training source-unit run recovered these 20 substantive gold rows:
+
+```text
+G5 G6 G7 G8 G9 G11 G13 G14 G15 G17 G20 G23 G25 G26 G31 G34 G37 G42 G44 G45
+```
+
+That result contains 16 direct rows and four derived rows. It gives a concrete
+bridge to the target:
+
+| Increment | New recoverable rows | Running total |
+|---|---|---:|
+| Current source-unit result | 20 reviewed matches | 20/46 |
+| Supply document metadata and run document-topic extraction | G1-G4 | 24/46 |
+| Licensed inverse/decomposition pass over accepted direct facts | G10, G12, G16, G18, G21, G24, G27, G32 | 32/46 |
+| Recover any five remaining causal/reified/discourse rows | Five of G19, G22, G28, G29, G30, G33, G35, G36, G38-G41, G43, G46 | **37/46** |
+| Declared margin target | At least one further row | **38/46** |
+
+The inverse increment is not a hard-coded predicate table. It is a trained,
+source-checked D1 generation task that emits a new semantic view only when the
+source plus accepted D0 fact licenses it. Free predicates remain free.
+
+This arithmetic is not a claimed future score. It demonstrates that the system
+does not have to solve all 26 remaining misses to reach 80%. It needs metadata,
+reliable learned inverse/decomposition behavior, and five to six improvements in
+the hard causal/reification/discourse set.
+
+## Exact target and evaluation contract
+
+### Gold taxonomy
+
+The machine-readable gold contains:
+
+| Tier | Rows | 80% floor | Plan gate |
+|---|---:|---:|---:|
+| Substantive direct (`h:false`, G1-G46) | 27 | 22 | 22 |
+| Substantive derived/interpretive (`h:true`, G1-G46) | 19 | 16 | 16 |
+| All substantive | 46 | 37 | **38** |
+| Structural `rdf:type` / `rdfs:label` | 36 | 29 | **29** |
+| Whole gold | 82 | 66 | **67** |
+
+### Matching rule
+
+Exact free-predicate string equality remains a conservative debugging signal, not
+the primary score. Evaluation uses one-to-one semantic fact families:
+
+- one generated fact may satisfy at most one gold row;
+- inverses count separately only where the gold explicitly includes the inverse;
+- subject/object direction, polarity, modality, temporal scope, attribution, and
+  important qualifiers must agree;
+- a vague whole-clause CURIE cannot match a more atomic gold proposition merely
+  because its spelling contains the right words;
+- unsupported facts count against precision even when another fact matches gold;
+- structural rows never contribute to the substantive score.
+
+This follows the fact-synset principle in BenchIE and avoids both exact-string
+under-credit and duplicate padding.
+
+### Admission gates
+
+The `llm_nerd` development fixture must satisfy all of:
+
+- substantive recall at least 38/46;
+- direct recall at least 22/27;
+- derived/interpretive recall at least 16/19;
+- structural recall at least 29/36 when structural mode is enabled;
+- unique supported-fact precision at least 90%;
+- no more than one clear role-direction or reflexive error;
+- zero malformed tool calls across three deterministic repeats;
+- bounded completion because all scheduled stages finish, not because the model
+  introspects global semantic exhaustion.
+
+The source-disjoint admission suite must satisfy:
+
+- mean substantive recall at least 80%;
+- lower 95% bootstrap confidence bound at least 75%;
+- no evaluation stratum below 70% recall;
+- mean unique supported-fact precision at least 90%;
+- role/direction error rate below 2%;
+- quantized RTX 3090 recall no more than three percentage points below the BF16
+  adapter result.
+
+## What the current model was actually trained to do
+
+The V3/V10 supervised pool is much narrower than the new gold contract:
+
+- 48 training source families and six validation families;
+- 881 training target facts and 79 validation facts;
+- the 881 training targets contain 600 reviewed semantic facts plus 281 generated
+  `rdfs:label` rows;
+- **zero** training targets have `h:true`;
+- **zero** training targets use `rdf:type`;
+- the training contract is direct, exact-anchor extraction with pagination and a
+  continuation flag;
+- later V5-V11 variants mostly reweighted or repartitioned these same source
+  families rather than adding new semantic supervision.
+
+The current adapter was therefore never given positive training examples for the
+19 `h:true` substantive rows in this gold, never taught type expansion, and was
+deployed behind a prompt that explicitly forbids inverse, derived, hypothesis,
+type, and label output. Expecting it to reach 80% of the new contract is a task
+mismatch, not merely undertraining.
+
+The old V10 H100 run processed roughly 9.85 million rendered token positions in
+5,731 seconds. Its low validation loss measured imitation of the narrow training
+serialization; it did not measure semantic coverage against the new gold.
+
+## Failure-to-component map
+
+### Currently recovered substantive rows
+
+| Gold rows | Capability already demonstrated |
+|---|---|
+| G5-G9 | belief, scope, and conditional decomposition |
+| G11 | evidence relation |
+| G13-G15, G17 | characterizations, counterparty relation, relative clause |
+| G20, G23, G25 | vendor/bank inability and geographic qualifier |
+| G26 | quoted leadership advocacy |
+| G31, G34, G37 | counter-China relation and alliance dependency |
+| G42, G44, G45 | contrast and author stance |
+| G6 | at least one useful inverse |
+
+### Missing substantive rows
+
+| Gold rows | Failure family | Planned recovery component |
+|---|---|---|
+| G1-G2 | author metadata was absent | metadata envelope plus metadata task |
+| G3-G4 | no document-as-entity topic pass | document-level D0/D1 pass |
+| G10, G12, G16, G18, G21, G24, G27, G32 | licensed inverses/decompositions omitted | D1 adapter task conditioned on accepted D0 ledger |
+| G19, G22 | consequence relation lost | source-unit D0 curriculum for causal markers and coordinated scope |
+| G28 | quote/coreference reference omitted | quotation plus coreference task |
+| G29-G30 | pragmatic implication plus inverse | separately labelled D2 task |
+| G33 | strategic event not reified | event/proposition reification curriculum |
+| G35-G36 | alternate alliance view and enabling relation | D1 relation-aware expansion |
+| G38-G39 | coordinated behavior entity not reified | coordination and event-mode examples |
+| G40-G41 | implied actor/target direction | D2 role inference plus validator |
+| G43 | cross-sentence strategic consequence | D2 discourse task |
+| G46 | argument-level conclusion | document-level D2 task |
+
+### Structural rows
+
+The 36 structural rows are not one undifferentiated problem:
+
+- 18 label rows come from an entity inventory and source/metadata display names;
+- 18 type rows require open-ended model classification;
+- their prerequisite is creating the correct entity, event, group, worldview,
+  effort, alliance, and approach nodes during D0/D1 extraction.
+
+The structural pass should consume the accepted entity inventory after semantic
+extraction. It should not compete with D0 fact generation for the same output page.
+
+## The target runtime architecture
+
+```text
+chunk text + document metadata
+    -> finite source-unit planner
+    -> parallel per-unit D0 extraction adapter
+    -> per-unit role/scope validator adapter
+    -> accepted D0 ledger
+    -> D1 licensed inverse/decomposition pass
+    -> optional D2 discourse/interpretation pass
+    -> entity type/label pass
+    -> semantic-family retrieval and deduplication
+    -> Donto ingest
+```
+
+### Stage 0: metadata envelope
+
+Always pass document IRI, author/creator when known, title, date, source kind, and
+the chunk's parent identifiers. The model may then emit document relations such as
+authorship and subject matter. Metadata-derived facts remain provenance-labelled
+so they cannot be mistaken for body-text assertions.
+
+### Stage 1: finite source traversal
+
+Partition the already-small Donto chunk into overlapping proposition-bearing
+units. Use stable source-relative IDs. The planner must cover:
+
+- finite verbs and predicative adjectives;
+- coordinated predicates and objects;
+- conditionals and consequences;
+- relative clauses and anaphora;
+- quotations and their referents;
+- contrastive branches;
+- document-level argument conclusions.
+
+The controller schedules every unit. Qwen's `more_supported_facts` bit does not
+control document completion. A final full-chunk residual pass checks for an
+unassigned proposition, but cannot create an unbounded loop.
+
+Within difficult units, the planner may also produce a dynamic entity/event-pair
+worklist and ask the extractor who did or was what to whom for each pair. This is
+relation-aware QA, not a fixed relation taxonomy. It applies the useful
+discriminative-to-generative idea while preserving Donto's free predicates.
+
+### Stage 2: D0 direct extraction
+
+Each unit is a short structured-output call. The generated internal record adds
+two fields that are stripped before public ingest:
+
+```json
+{
+  "sourceUnitId": "u04",
+  "layer": "D0",
+  "s": "ex:donald-trump",
+  "p": "unableToFind",
+  "o": "ex:bank-outside-russia",
+  "c": 0.98,
+  "h": false
+}
+```
+
+Train the model to split conjunctions, preserve qualifiers, create event or
+proposition entities only when necessary, and read the relation back from subject
+to object before submission.
+
+### Stage 3: learned validator/repairer
+
+For each candidate, provide the source unit, resolved local context, and candidate
+fact. Force one of:
+
+```text
+ACCEPT | REVERSE | REWRITE | REJECT | AMBIGUOUS
+```
+
+`REVERSE` and `REWRITE` require a complete replacement fact. This targets the
+five clear hosted errors without asking one broad verifier to regenerate the whole
+ledger. The no-training experiment showed that broad verification either failed
+its tool call or over-pruned good facts.
+
+### Stage 4: D1 semantic views
+
+Condition on the accepted D0 ledger and source. Generate useful new, source-
+licensed inverses, decompositions, evidential links, and event relations. Mark
+every row `h:true`. Do not use a hand-maintained inverse-predicate dictionary.
+
+### Stage 5: D2 discourse interpretations
+
+Run a separate bounded task for stance, irony, contradictions, likely omitted
+roles, and cross-sentence consequences. D2 is allowed to abstain. Its facts are
+`h:true` and separately measurable. It never prolongs D0 traversal.
+
+### Stage 6: structural expansion
+
+Given the accepted entity inventory, emit one label and, where supported, one
+open-ended type per entity. Labels should prefer source or metadata wording. This
+stage is separately switchable and cannot inflate substantive recall.
+
+### Stage 7: semantic aggregation
+
+Use the existing Donto embeddings/alignment system to retrieve candidate
+equivalents. A learned source-aware classifier decides semantic family,
+contradiction, and novelty. Store equivalent surface variants in audit provenance
+rather than deleting them. This follows Donto's emit-free/defer-joining model and
+does not introduce a predicate synonym table.
+
+## Training corpus plan
+
+### Splits and scale
+
+Do not train on the `llm_nerd` source, close paraphrases of it, its gold, or model
+outputs labelled against it. Keep it as a development fixture.
+
+Build in two rounds:
+
+| Round | Training sources | Validation sources | Purpose |
+|---|---:|---:|---|
+| Pilot | existing 48 plus at least 100 new | 30 new | establish whether the architecture reaches 65-70% source-disjoint recall |
+| Main | 400-500 total | 60 new | reach and stabilize the 80% target |
+
+Preserve a separate sealed 100-source blind suite. Source families, near
+duplicates, document parents, and synthetic templates must be disjoint across all
+splits.
+
+The main corpus should target approximately:
+
+- 15,000-25,000 accepted substantive facts;
+- 8,000-15,000 D0 unit-generation rows;
+- 4,000-8,000 D1/D2 and structural rows;
+- 15,000-30,000 validator decisions, including Qwen-induced mistakes.
+
+These are planning ranges. Progress is decided by semantic learning curves, not
+by reaching a decorative row count.
+
+### Source diversity
+
+Stratify real Donto chunks across argumentative comments, news/narrative prose,
+historical OCR, technical/log material, academic analysis, and
+genealogical/administrative records. Within every stratum deliberately sample:
+
+- conditionals, negation, modality, attribution, quotation, and reported speech;
+- coordination and shared modifiers;
+- relative clauses and pronouns;
+- causal/evidential markers;
+- temporal and geographic qualifiers;
+- evaluative language and cross-sentence conclusions.
+
+Source-balanced batching prevents a high-volume source or repeated predicate
+family from dominating the gradient.
+
+### Teacher construction
+
+Use Codex SOTA with the canonical Donto abundance contract as the teacher, in
+separate source-first passes:
+
+1. source-unit plan;
+2. D0 direct gold;
+3. D1 licensed views;
+4. D2 interpretations;
+5. entity type/label inventory;
+6. independent source-bound validation and repair.
+
+Retain the teacher prompt, rendered input, tool schema, raw output, adjudication,
+source hash, model/version, and accepted/rejected IDs. Human-review a stratified
+sample and every disputed or high-impact row. Do not use the OpenRouter-hosted
+Qwen output as teacher truth; use it to generate realistic student mistakes.
+
+### Synthetic data
+
+Use structure-to-text generation aggressively when reviewed real families are
+the limiting resource, while keeping the counts explicit:
+
+1. specify a small target graph with roles, qualifiers, inverse views, and layer;
+2. ask the teacher to write natural text that expresses it;
+3. independently re-extract from the text;
+4. retain only facts supported by both the source-first review and graph contract.
+
+This is valuable for rare role/scope patterns because generating text from a
+known structure is easier to audit than inventing labels for arbitrary text.
+Synthetic and real data remain separately counted and reported.
+
+### On-policy hard negatives
+
+After each candidate adapter:
+
+1. run it on training sources only;
+2. capture role reversals, reflexive edges, dropped coordination, scope leakage,
+   opaque clause objects, unsupported causality, duplicates, and tool-call
+   failures;
+3. have the source-bound teacher label or repair those exact visited states;
+4. aggregate them into the next validator/extractor corpus.
+
+This DAgger-like loop fixes the distribution gap between pristine teacher
+histories and the states created by Qwen's own prior outputs.
+
+## Training recipe
+
+### Coverage adapter
+
+Start every candidate from the pinned stock Qwen3.8-27B revision. Never continue
+from V3-V11.
+
+Recommended first recipe:
+
+- QLoRA with NF4, double quantization, and BF16 compute;
+- 8,192-token maximum after measuring the new rendered-length distribution; no
+  truncation or dropped rows;
+- completion/assistant-only loss;
+- source-balanced, length-bucketed batches;
+- one epoch initially, cosine schedule, 3-5% warmup;
+- learning-rate ablation at `1e-5`, `2e-5`, and `5e-5` on the pilot;
+- compare rank-16 conventional LoRA against rank-32 rsLoRA;
+- dropout 0.05;
+- semantic recall/precision checkpoint selection, not aggregate validation loss.
+
+The final learning rate and rank are deliberately not declared in advance. The
+old 5e-5/rank-16 recipe optimized a different task and adapted an incomplete
+module set.
+
+### Correct hybrid-module targets
+
+The existing V10 adapter missed every Gated DeltaNet input projection. The
+patched trainer now defines the required 496 text-only modules:
+
+| Family | Count |
+|---|---:|
+| Full-attention `q_proj/k_proj/v_proj/o_proj` | 64 |
+| DeltaNet `in_proj_qkv/in_proj_z/in_proj_a/in_proj_b/out_proj` | 240 |
+| MLP `gate_proj/up_proj/down_proj` | 192 |
+| Vision modules | 0 |
+
+Every run must persist the resolved module and parameter names and fail before
+training if the census differs. PEFT's QLoRA guidance recommends adapting all
+linear layers; the language-backbone regex achieves that without touching the
+vision tower.
+
+### Validator adapter
+
+Train a separate rank-16 adapter on the five-way validation/repair task. A
+separate adapter prevents extraction abundance from competing directly with
+conservative rejection. vLLM can serve multiple LoRA adapters over one base.
+
+Begin with structured SFT. After SFT, optionally test a short DPO phase using
+accepted/repaired output as chosen and the student's defective output as rejected:
+
+- start around `1e-6` learning rate and `beta=0.1`;
+- keep the reference adapter frozen;
+- run no more than a partial epoch before evaluation;
+- retain DPO only if it improves source-disjoint precision without reducing
+  recall or tool-call reliability.
+
+DPO is an ablation, not an article of faith. Structured repair SFT may be enough.
+
+### Task mixture
+
+Use explicit task contracts rather than mixing every behavior in one unlabelled
+assistant distribution:
+
+| Task | Initial sampling weight |
+|---|---:|
+| Source-unit D0 extraction | 45% |
+| D1 inverse/decomposition | 20% |
+| D2 discourse interpretation | 10% |
+| Metadata/topic extraction | 5% |
+| Entity type/label expansion | 10% |
+| Residual coverage probe | 10% |
+
+The validator corpus is trained separately. Rebalance these weights using per-task
+validation metrics; do not replicate the same terminal row dozens of times.
+
+### What not to train again
+
+- no global `more_supported_facts` objective as the completion authority;
+- no 18K raw multi-page histories merely to teach stopping;
+- no anchor copying objective while anchors are out of scope;
+- no type/label facts mixed into the same capped page as substantive D0 facts;
+- no rejected adapter as the starting weights;
+- no evaluation fixture or close paraphrase in training;
+- no exact predicate vocabulary or inverse lookup table.
+
+## Experiment ladder
+
+Run the following sequentially, stopping a branch as soon as it fails its gate:
+
+| Experiment | Change from previous | Required result |
+|---|---|---|
+| E0 | Frozen current local unit pipeline | Reproduce baseline and manual score |
+| E1 | New controller, metadata, D1/D2/structural stages; no retraining | Establish architectural ceiling and error corpus |
+| E2 | Pilot D0/D1 coverage SFT, corrected 496-module target set | At least 65% substantive recall on new development sources; precision at least 90% |
+| E3 | Add on-policy validator SFT | Halve role/scope error rate without losing more than two recall points |
+| E4 | Add D2 and structural curriculum | `llm_nerd` at least 38/46 and 67/82; new-suite mean at least 75% |
+| E5 | Scale to 400-500 source families | New-suite mean at least 80% and all admission gates |
+| E6 | Rank/LR/seed confirmation | Same recipe passes with at least two of three seeds |
+| E7 | W4A16 RTX 3090 deployment | No more than three recall points below BF16 and runtime stability passes |
+
+Required ablations:
+
+- corrected DeltaNet inputs versus the old incomplete target set;
+- rank 16 versus rank 32 rsLoRA;
+- gold-only training versus gold plus on-policy hard negatives;
+- one combined adapter versus separate validator adapter;
+- D0 only versus D0+D1 versus D0+D1+D2;
+- BF16 adapter versus runtime LoRA over the exact W4A16 base;
+- prefix cache/MTP enabled versus disabled for quality parity;
+- 4K versus 8K training context after length measurement.
+
+## Evaluation suite
+
+### Development
+
+- retain the current six sources as deterministic regression fixtures;
+- retain `llm_nerd` as the explicit 80%-gold development target;
+- add 60 source-disjoint development documents with complete fact-synset gold;
+- report direct, D1, D2, structural, precision, recall, duplicate rate, role errors,
+  tool-call failures, and calls after exhaustion separately.
+
+### Blind
+
+- seal 100 source-disjoint documents before main training;
+- include at least ten documents from each important source/failure stratum;
+- evaluate complete trajectories, not a sample of isolated emitted facts;
+- adjudicate identity-blind against the stock model, V10, the new BF16 adapter,
+  and the quantized RTX 3090 system;
+- keep every miss, false positive, contradiction, and reviewer disagreement.
+
+### Learning curves
+
+At 100, 200, 300, 400, and 500 training source families, evaluate the same frozen
+development suite. Stop adding data only when the curve flattens and the 80% gate
+is stable across seeds. The number of duplicated rows is not a substitute for
+independent sources.
+
+## Hardware, time, and cost
+
+Train and serve only on the operator-selected RunPod RTX 3090. The rendered
+training rows top out at 2,263 tokens for coverage and 2,562 for the validator,
+so 4K retains the entire corpus. Standard full-vocabulary cross-entropy exceeded
+24GB on the longest row; Liger fused loss completed the maximum-row smoke at a
+safe working set. Training cost is therefore budgeted from measured optimizer
+step time on this exact card, not from the old H100 run or raw row count.
+
+Planning envelope:
+
+| Work | Measured/projected RTX 3090 time |
+|---|---:|
+| Pilot coverage adapter | about 1.5-2 hours including validation/save |
+| Pilot validator adapter | about 1 hour including validation/save |
+| Main coverage candidate per seed | about 6 hours plus semantic evaluation |
+| Main validator candidate | about 4 hours plus semantic evaluation |
+| Two-of-three seed confirmation | run sequentially; approximately 12-18 hours for coverage |
+
+These are measured projections, not guarantees or price quotes. Do not start a
+run until the immutable dataset, rendered-length audit, run contract, fused-loss
+maximum-row smoke, and storage floor pass.
+
+Serve the admitted system first on the RTX 3090. The 8GB RTX 3070 remains an
+offload experiment, not the quality target. The `syv-ai/qwen38-27b-rtx3090`
+runtime optimizations are useful after semantic admission:
+
+- keep prefix caching because every unit call shares the same instruction,
+  schema, document, and ledger prefix;
+- retain exact speculative decoding/MTP only after parity tests;
+- use the short-context profile because Donto already chunks documents;
+- batch independent units from one or several documents;
+- do not enable more aggressive lossy activation or cache quantization until the
+  80% suite proves no material regression.
+
+If runtime LoRA over the exact W4 base loses more than three recall points against
+BF16, merge into a high-precision base, requantize using representative structured
+extraction trajectories, and rerun every semantic gate.
+
+## Operational safeguards
+
+- Keep the production extraction lane paused through E0-E6.
+- Use a separate opt-in `qwen-quality-canary` lane for E7.
+- Never expose blind targets to the trainer, teacher, prompt designer, or
+  candidate selector.
+- Hash every source, split, prompt, schema, rendered fixture, model revision,
+  adapter, quantization artifact, evaluation packet, and score file.
+- Train fresh from stock for every candidate.
+- Persist failures and rejected adapters; never silently promote the best-looking
+  single seed.
+- Verify actual GPU progress, output growth, RAM/swap, and billed pod lifetime.
+- Stop and terminate the RTX 3090 pod immediately after durable artifacts are copied and
+  hash-verified.
+
+## Recommended execution order
+
+1. Freeze the 46-row substantive and 82-row whole-gold match ledgers.
+2. Implement the finite multi-stage controller behind a non-production flag.
+3. Run E1 with current Qwen locally and capture exact on-policy errors.
+4. Build and audit the 100-new-source pilot corpus.
+5. Run the corrected module-target/rank/LR pilot ablations from stock.
+6. Train the validator on the pilot model's own mistakes.
+7. Evaluate E2-E4 on `llm_nerd` and 60 new development sources.
+8. If the learning curve supports the target, expand to 400-500 training sources.
+9. Train three final seeds, select by source-disjoint semantic metrics, and open
+   the blind packet once.
+10. Quantize/serve on the RTX 3090 canary and compare complete trajectories.
+11. Promote only after all semantic and runtime gates pass.
+
+## The plan in one sentence
+
+Train Qwen to exhaust short source units and generate separately labelled D0/D1/D2
+views, train a second Qwen adapter on its own role/scope mistakes, and let a finite
+controller assemble the result; this is the most credible route from the measured
+20/46 baseline to at least 38/46 without turning Donto into a fixed-ontology
+extractor.
+
+## Primary research and implementation sources
+
+- Qwen3.8 model and reasoning/tool-use contract:
+  https://huggingface.co/Qwen/Qwen3.8-27B
+- PEFT LoRA targeting and QLoRA guidance:
+  https://huggingface.co/docs/peft/package_reference/lora
+- QLoRA:
+  https://arxiv.org/abs/2305.14314
+- Rank-stabilized LoRA:
+  https://arxiv.org/abs/2312.03732
+- TRL assistant/completion-only SFT:
+  https://huggingface.co/docs/trl/sft_trainer
+- DPO and TRL implementation:
+  https://arxiv.org/abs/2305.18290
+  and https://huggingface.co/docs/trl/en/dpo_trainer
+- Iterative extraction conditioned on prior results (IMoJIE):
+  https://aclanthology.org/2020.acl-main.521/
+- QA-aligned relation extraction (QA4RE):
+  https://aclanthology.org/2023.findings-acl.50/
+- Relation-aware two-stage extraction:
+  https://aclanthology.org/2025.starsem-1.22/
+- Discriminative-to-generative candidate guidance:
+  https://aclanthology.org/2026.acl-long.2093/
+- Discovery, denoising, and prediction for open relations:
+  https://aclanthology.org/2025.emnlp-main.459/
+- Fact-synset evaluation (BenchIE):
+  https://aclanthology.org/2022.acl-long.307/
+- Synthetic structure-to-text information extraction:
+  https://arxiv.org/abs/2303.04132
+- On-policy dataset aggregation:
+  https://proceedings.mlr.press/v15/ross11a.html
+- vLLM multi-adapter serving:
+  https://docs.vllm.ai/en/stable/features/lora/
+- RTX 3090 Qwen3.8 serving and measured optimization tradeoffs:
+  https://github.com/syv-ai/qwen38-27b-rtx3090
+- Current RunPod pricing:
+  https://www.runpod.io/pricing

docs/UPSTREAM-ADAPTER-README.md

@@ -0,0 +1,207 @@
+---
+base_model: /workspace/models/Qwen3.8-27B-BF16
+library_name: peft
+pipeline_tag: text-generation
+tags:
+- base_model:adapter:/workspace/models/Qwen3.8-27B-BF16
+- lora
+- transformers
+---
+
+# Model Card for Model ID
+
+<!-- Provide a quick summary of what the model is/does. -->
+
+
+
+## Model Details
+
+### Model Description
+
+<!-- Provide a longer summary of what this model is. -->
+
+
+
+- **Developed by:** [More Information Needed]
+- **Funded by [optional]:** [More Information Needed]
+- **Shared by [optional]:** [More Information Needed]
+- **Model type:** [More Information Needed]
+- **Language(s) (NLP):** [More Information Needed]
+- **License:** [More Information Needed]
+- **Finetuned from model [optional]:** [More Information Needed]
+
+### Model Sources [optional]
+
+<!-- Provide the basic links for the model. -->
+
+- **Repository:** [More Information Needed]
+- **Paper [optional]:** [More Information Needed]
+- **Demo [optional]:** [More Information Needed]
+
+## Uses
+
+<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
+
+### Direct Use
+
+<!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->
+
+[More Information Needed]
+
+### Downstream Use [optional]
+
+<!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->
+
+[More Information Needed]
+
+### Out-of-Scope Use
+
+<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->
+
+[More Information Needed]
+
+## Bias, Risks, and Limitations
+
+<!-- This section is meant to convey both technical and sociotechnical limitations. -->
+
+[More Information Needed]
+
+### Recommendations
+
+<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->
+
+Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.
+
+## How to Get Started with the Model
+
+Use the code below to get started with the model.
+
+[More Information Needed]
+
+## Training Details
+
+### Training Data
+
+<!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->
+
+[More Information Needed]
+
+### Training Procedure
+
+<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->
+
+#### Preprocessing [optional]
+
+[More Information Needed]
+
+
+#### Training Hyperparameters
+
+- **Training regime:** [More Information Needed] <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
+
+#### Speeds, Sizes, Times [optional]
+
+<!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->
+
+[More Information Needed]
+
+## Evaluation
+
+<!-- This section describes the evaluation protocols and provides the results. -->
+
+### Testing Data, Factors & Metrics
+
+#### Testing Data
+
+<!-- This should link to a Dataset Card if possible. -->
+
+[More Information Needed]
+
+#### Factors
+
+<!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->
+
+[More Information Needed]
+
+#### Metrics
+
+<!-- These are the evaluation metrics being used, ideally with a description of why. -->
+
+[More Information Needed]
+
+### Results
+
+[More Information Needed]
+
+#### Summary
+
+
+
+## Model Examination [optional]
+
+<!-- Relevant interpretability work for the model goes here -->
+
+[More Information Needed]
+
+## Environmental Impact
+
+<!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->
+
+Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700).
+
+- **Hardware Type:** [More Information Needed]
+- **Hours used:** [More Information Needed]
+- **Cloud Provider:** [More Information Needed]
+- **Compute Region:** [More Information Needed]
+- **Carbon Emitted:** [More Information Needed]
+
+## Technical Specifications [optional]
+
+### Model Architecture and Objective
+
+[More Information Needed]
+
+### Compute Infrastructure
+
+[More Information Needed]
+
+#### Hardware
+
+[More Information Needed]
+
+#### Software
+
+[More Information Needed]
+
+## Citation [optional]
+
+<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
+
+**BibTeX:**
+
+[More Information Needed]
+
+**APA:**
+
+[More Information Needed]
+
+## Glossary [optional]
+
+<!-- If relevant, include terms and calculations in this section that can help readers understand the model or model card. -->
+
+[More Information Needed]
+
+## More Information [optional]
+
+[More Information Needed]
+
+## Model Card Authors [optional]
+
+[More Information Needed]
+
+## Model Card Contact
+
+[More Information Needed]
+### Framework versions
+
+- PEFT 0.20.0

evaluation/pilot-live-v15-exhaustive-unit-r16-v1-validation-canary1.json

@@ -0,0 +1,173 @@
+{
+  "aggregate": {
+    "counts": {
+      "calls_failed": 0,
+      "calls_ok": 1,
+      "duplicate_items": 0,
+      "invalid_predicted_items": 0,
+      "matched_items": 2,
+      "predicted_items": 3,
+      "reflexive_entity_items": 0,
+      "rows": 1,
+      "target_items": 3,
+      "valid_predicted_items": 3
+    },
+    "exact_precision": 0.6666666666666666,
+    "exact_recall": 0.6666666666666666,
+    "tool_call_success_rate": 1.0
+  },
+  "bootstrap_95pct_lower_bound": 0.6666666666666666,
+  "by_document": {
+    "rail-incident-000": {
+      "counts": {
+        "calls_failed": 0,
+        "calls_ok": 1,
+        "duplicate_items": 0,
+        "invalid_predicted_items": 0,
+        "matched_items": 2,
+        "predicted_items": 3,
+        "reflexive_entity_items": 0,
+        "rows": 1,
+        "target_items": 3,
+        "valid_predicted_items": 3
+      },
+      "exact_precision": 0.6666666666666666,
+      "exact_recall": 0.6666666666666666,
+      "tool_call_success_rate": 1.0
+    }
+  },
+  "by_group": {
+    "substantive": {
+      "counts": {
+        "calls_failed": 0,
+        "calls_ok": 1,
+        "duplicate_items": 0,
+        "invalid_predicted_items": 0,
+        "matched_items": 2,
+        "predicted_items": 3,
+        "reflexive_entity_items": 0,
+        "rows": 1,
+        "target_items": 3,
+        "valid_predicted_items": 3
+      },
+      "exact_precision": 0.6666666666666666,
+      "exact_recall": 0.6666666666666666,
+      "tool_call_success_rate": 1.0
+    }
+  },
+  "by_stratum": {
+    "incident-rail-style0": {
+      "counts": {
+        "calls_failed": 0,
+        "calls_ok": 1,
+        "duplicate_items": 0,
+        "invalid_predicted_items": 0,
+        "matched_items": 2,
+        "predicted_items": 3,
+        "reflexive_entity_items": 0,
+        "rows": 1,
+        "target_items": 3,
+        "valid_predicted_items": 3
+      },
+      "exact_precision": 0.6666666666666666,
+      "exact_recall": 0.6666666666666666,
+      "tool_call_success_rate": 1.0
+    }
+  },
+  "by_task": {
+    "metadata": {
+      "counts": {
+        "calls_failed": 0,
+        "calls_ok": 1,
+        "duplicate_items": 0,
+        "invalid_predicted_items": 0,
+        "matched_items": 2,
+        "predicted_items": 3,
+        "reflexive_entity_items": 0,
+        "rows": 1,
+        "target_items": 3,
+        "valid_predicted_items": 3
+      },
+      "exact_precision": 0.6666666666666666,
+      "exact_recall": 0.6666666666666666,
+      "tool_call_success_rate": 1.0
+    }
+  },
+  "manifest": {
+    "args": {
+      "allow_blind": false,
+      "concurrency": 1,
+      "dataset": "/mnt/donto-data/donto-resources/extraction/qwen38-distill-20260818/80-percent-gold-plan-20260822/pilot-training-live-v15-exhaustive-unit-v3/validation.jsonl",
+      "limit": 1,
+      "max_tokens": 1024,
+      "model": "pilot-live-v15-exhaustive-unit-r16-v1",
+      "output": "/mnt/donto-data/donto-resources/extraction/qwen38-distill-20260818/80-percent-gold-plan-20260822/results/pilot-live-v15-exhaustive-unit-r16-v1-validation-canary1.json",
+      "provider": "runpod",
+      "tasks": [],
+      "timeout_seconds": 300,
+      "tools": "/mnt/donto-data/donto-resources/extraction/qwen38-distill-20260818/80-percent-gold-plan-20260822/pilot-training-live-v15-exhaustive-unit-v3/tools.json"
+    },
+    "binary_sha256": "b0e7092557edd51a3fd5e22edc61649319cc010b07c9e8e9d92a93988e9a4f03",
+    "blind_holdout_opened": false,
+    "created_at": "2026-08-25T09:20:52.126747637+00:00",
+    "dataset": "/mnt/donto-data/donto-resources/extraction/qwen38-distill-20260818/80-percent-gold-plan-20260822/pilot-training-live-v15-exhaustive-unit-v3/validation.jsonl",
+    "dataset_sha256": "c360e8296e5ec7973f2fab8416d0b30e2aa850fe6b2cc618bb3fa52894d818b2",
+    "elapsed_ms": 9603,
+    "git_hash": "12ab526",
+    "model": "pilot-live-v15-exhaustive-unit-r16-v1",
+    "provider": "runpod",
+    "schema": "donto-qwen38-structured-corpus-eval-v1",
+    "tool_name": "submit_facts",
+    "tools": "/mnt/donto-data/donto-resources/extraction/qwen38-distill-20260818/80-percent-gold-plan-20260822/pilot-training-live-v15-exhaustive-unit-v3/tools.json",
+    "tools_sha256": "e02cf3c20831eed00a05598f42f324f9112f7c283b9cd0cef43af96e294c93ab"
+  },
+  "mean_document_exact_recall": 0.6666666666666666,
+  "rows": [
+    {
+      "arguments": {
+        "facts": [
+          {
+            "c": 0.99,
+            "h": false,
+            "o": "ex:ari-alder-000",
+            "p": "createdBy",
+            "s": "ex:rail-operations-review-000"
+          },
+          {
+            "c": 0.99,
+            "h": false,
+            "o": "2032-01-01",
+            "p": "dated",
+            "s": "ex:rail-operations-review-000"
+          },
+          {
+            "c": 0.99,
+            "h": false,
+            "o": "ex:rail-review-panel-000",
+            "p": "discusses",
+            "s": "ex:rail-operations-review-000"
+          }
+        ],
+        "more_supported_facts": false
+      },
+      "counts": {
+        "calls_failed": 0,
+        "calls_ok": 1,
+        "duplicate_items": 0,
+        "invalid_predicted_items": 0,
+        "matched_items": 2,
+        "predicted_items": 3,
+        "reflexive_entity_items": 0,
+        "rows": 1,
+        "target_items": 3,
+        "valid_predicted_items": 3
+      },
+      "document_id": "rail-incident-000",
+      "elapsed_ms": 9435,
+      "error": null,
+      "index": 0,
+      "parent_template": "incident-rail-style0",
+      "task": "metadata"
+    }
+  ]
+}

evaluation/pilot-live-v15-exhaustive-unit-r16-v1-validation-first30.json

@@ -0,0 +1,1414 @@
+{
+  "aggregate": {
+    "counts": {
+      "calls_failed": 0,
+      "calls_ok": 30,
+      "duplicate_items": 0,
+      "invalid_predicted_items": 0,
+      "matched_items": 52,
+      "predicted_items": 57,
+      "reflexive_entity_items": 0,
+      "rows": 30,
+      "target_items": 57,
+      "valid_predicted_items": 57
+    },
+    "exact_precision": 0.9122807017543859,
+    "exact_recall": 0.9122807017543859,
+    "tool_call_success_rate": 1.0
+  },
+  "bootstrap_95pct_lower_bound": 0.8913043478260869,
+  "by_document": {
+    "rail-incident-000": {
+      "counts": {
+        "calls_failed": 0,
+        "calls_ok": 26,
+        "duplicate_items": 0,
+        "invalid_predicted_items": 0,
+        "matched_items": 41,
+        "predicted_items": 46,
+        "reflexive_entity_items": 0,
+        "rows": 26,
+        "target_items": 46,
+        "valid_predicted_items": 46
+      },
+      "exact_precision": 0.8913043478260869,
+      "exact_recall": 0.8913043478260869,
+      "tool_call_success_rate": 1.0
+    },
+    "university-quote-001": {
+      "counts": {
+        "calls_failed": 0,
+        "calls_ok": 4,
+        "duplicate_items": 0,
+        "invalid_predicted_items": 0,
+        "matched_items": 11,
+        "predicted_items": 11,
+        "reflexive_entity_items": 0,
+        "rows": 4,
+        "target_items": 11,
+        "valid_predicted_items": 11
+      },
+      "exact_precision": 1.0,
+      "exact_recall": 1.0,
+      "tool_call_success_rate": 1.0
+    }
+  },
+  "by_group": {
+    "substantive": {
+      "counts": {
+        "calls_failed": 0,
+        "calls_ok": 30,
+        "duplicate_items": 0,
+        "invalid_predicted_items": 0,
+        "matched_items": 52,
+        "predicted_items": 57,
+        "reflexive_entity_items": 0,
+        "rows": 30,
+        "target_items": 57,
+        "valid_predicted_items": 57
+      },
+      "exact_precision": 0.9122807017543859,
+      "exact_recall": 0.9122807017543859,
+      "tool_call_success_rate": 1.0
+    }
+  },
+  "by_stratum": {
+    "incident-rail-style0": {
+      "counts": {
+        "calls_failed": 0,
+        "calls_ok": 26,
+        "duplicate_items": 0,
+        "invalid_predicted_items": 0,
+        "matched_items": 41,
+        "predicted_items": 46,
+        "reflexive_entity_items": 0,
+        "rows": 26,
+        "target_items": 46,
+        "valid_predicted_items": 46
+      },
+      "exact_precision": 0.8913043478260869,
+      "exact_recall": 0.8913043478260869,
+      "tool_call_success_rate": 1.0
+    },
+    "quote-university-style0": {
+      "counts": {
+        "calls_failed": 0,
+        "calls_ok": 4,
+        "duplicate_items": 0,
+        "invalid_predicted_items": 0,
+        "matched_items": 11,
+        "predicted_items": 11,
+        "reflexive_entity_items": 0,
+        "rows": 4,
+        "target_items": 11,
+        "valid_predicted_items": 11
+      },
+      "exact_precision": 1.0,
+      "exact_recall": 1.0,
+      "tool_call_success_rate": 1.0
+    }
+  },
+  "by_task": {
+    "d0": {
+      "counts": {
+        "calls_failed": 0,
+        "calls_ok": 7,
+        "duplicate_items": 0,
+        "invalid_predicted_items": 0,
+        "matched_items": 24,
+        "predicted_items": 28,
+        "reflexive_entity_items": 0,
+        "rows": 7,
+        "target_items": 28,
+        "valid_predicted_items": 28
+      },
+      "exact_precision": 0.8571428571428571,
+      "exact_recall": 0.8571428571428571,
+      "tool_call_success_rate": 1.0
+    },
+    "d1": {
+      "counts": {
+        "calls_failed": 0,
+        "calls_ok": 1,
+        "duplicate_items": 0,
+        "invalid_predicted_items": 0,
+        "matched_items": 0,
+        "predicted_items": 0,
+        "reflexive_entity_items": 0,
+        "rows": 1,
+        "target_items": 0,
+        "valid_predicted_items": 0
+      },
+      "exact_precision": null,
+      "exact_recall": null,
+      "tool_call_success_rate": 1.0
+    },
+    "d1_unit": {
+      "counts": {
+        "calls_failed": 0,
+        "calls_ok": 8,
+        "duplicate_items": 0,
+        "invalid_predicted_items": 0,
+        "matched_items": 5,
+        "predicted_items": 5,
+        "reflexive_entity_items": 0,
+        "rows": 8,
+        "target_items": 5,
+        "valid_predicted_items": 5
+      },
+      "exact_precision": 1.0,
+      "exact_recall": 1.0,
+      "tool_call_success_rate": 1.0
+    },
+    "d2": {
+      "counts": {
+        "calls_failed": 0,
+        "calls_ok": 2,
+        "duplicate_items": 0,
+        "invalid_predicted_items": 0,
+        "matched_items": 3,
+        "predicted_items": 3,
+        "reflexive_entity_items": 0,
+        "rows": 2,
+        "target_items": 3,
+        "valid_predicted_items": 3
+      },
+      "exact_precision": 1.0,
+      "exact_recall": 1.0,
+      "tool_call_success_rate": 1.0
+    },
+    "metadata": {
+      "counts": {
+        "calls_failed": 0,
+        "calls_ok": 2,
+        "duplicate_items": 0,
+        "invalid_predicted_items": 0,
+        "matched_items": 5,
+        "predicted_items": 6,
+        "reflexive_entity_items": 0,
+        "rows": 2,
+        "target_items": 6,
+        "valid_predicted_items": 6
+      },
+      "exact_precision": 0.8333333333333334,
+      "exact_recall": 0.8333333333333334,
+      "tool_call_success_rate": 1.0
+    },
+    "residual": {
+      "counts": {
+        "calls_failed": 0,
+        "calls_ok": 2,
+        "duplicate_items": 0,
+        "invalid_predicted_items": 0,
+        "matched_items": 3,
+        "predicted_items": 3,
+        "reflexive_entity_items": 0,
+        "rows": 2,
+        "target_items": 3,
+        "valid_predicted_items": 3
+      },
+      "exact_precision": 1.0,
+      "exact_recall": 1.0,
+      "tool_call_success_rate": 1.0
+    },
+    "residual_unit": {
+      "counts": {
+        "calls_failed": 0,
+        "calls_ok": 6,
+        "duplicate_items": 0,
+        "invalid_predicted_items": 0,
+        "matched_items": 0,
+        "predicted_items": 0,
+        "reflexive_entity_items": 0,
+        "rows": 6,
+        "target_items": 0,
+        "valid_predicted_items": 0
+      },
+      "exact_precision": null,
+      "exact_recall": null,
+      "tool_call_success_rate": 1.0
+    },
+    "structural_batch": {
+      "counts": {
+        "calls_failed": 0,
+        "calls_ok": 2,
+        "duplicate_items": 0,
+        "invalid_predicted_items": 0,
+        "matched_items": 12,
+        "predicted_items": 12,
+        "reflexive_entity_items": 0,
+        "rows": 2,
+        "target_items": 12,
+        "valid_predicted_items": 12
+      },
+      "exact_precision": 1.0,
+      "exact_recall": 1.0,
+      "tool_call_success_rate": 1.0
+    }
+  },
+  "manifest": {
+    "args": {
+      "allow_blind": false,
+      "concurrency": 4,
+      "dataset": "/mnt/donto-data/donto-resources/extraction/qwen38-distill-20260818/80-percent-gold-plan-20260822/pilot-training-live-v15-exhaustive-unit-v3/validation.jsonl",
+      "limit": 30,
+      "max_tokens": 1024,
+      "model": "pilot-live-v15-exhaustive-unit-r16-v1",
+      "output": "/mnt/donto-data/donto-resources/extraction/qwen38-distill-20260818/80-percent-gold-plan-20260822/results/pilot-live-v15-exhaustive-unit-r16-v1-validation-first30.json",
+      "provider": "runpod",
+      "tasks": [],
+      "timeout_seconds": 300,
+      "tools": "/mnt/donto-data/donto-resources/extraction/qwen38-distill-20260818/80-percent-gold-plan-20260822/pilot-training-live-v15-exhaustive-unit-v3/tools.json"
+    },
+    "binary_sha256": "b0e7092557edd51a3fd5e22edc61649319cc010b07c9e8e9d92a93988e9a4f03",
+    "blind_holdout_opened": false,
+    "created_at": "2026-08-25T09:22:37.136692327+00:00",
+    "dataset": "/mnt/donto-data/donto-resources/extraction/qwen38-distill-20260818/80-percent-gold-plan-20260822/pilot-training-live-v15-exhaustive-unit-v3/validation.jsonl",
+    "dataset_sha256": "c360e8296e5ec7973f2fab8416d0b30e2aa850fe6b2cc618bb3fa52894d818b2",
+    "elapsed_ms": 94457,
+    "git_hash": "12ab526",
+    "model": "pilot-live-v15-exhaustive-unit-r16-v1",
+    "provider": "runpod",
+    "schema": "donto-qwen38-structured-corpus-eval-v1",
+    "tool_name": "submit_facts",
+    "tools": "/mnt/donto-data/donto-resources/extraction/qwen38-distill-20260818/80-percent-gold-plan-20260822/pilot-training-live-v15-exhaustive-unit-v3/tools.json",
+    "tools_sha256": "e02cf3c20831eed00a05598f42f324f9112f7c283b9cd0cef43af96e294c93ab"
+  },
+  "mean_document_exact_recall": 0.9456521739130435,
+  "rows": [
+    {
+      "arguments": {
+        "facts": [
+          {
+            "c": 0.99,
+            "h": false,
+            "o": "ex:ari-alder-000",
+            "p": "createdBy",
+            "s": "ex:rail-operations-review-000"
+          },
+          {
+            "c": 0.99,
+            "h": false,
+            "o": "2032-01-01",
+            "p": "dated",
+            "s": "ex:rail-operations-review-000"
+          },
+          {
+            "c": 0.99,
+            "h": false,
+            "o": "ex:rail-review-panel-000",
+            "p": "discusses",
+            "s": "ex:rail-operations-review-000"
+          }
+        ],
+        "more_supported_facts": false
+      },
+      "counts": {
+        "calls_failed": 0,
+        "calls_ok": 1,
+        "duplicate_items": 0,
+        "invalid_predicted_items": 0,
+        "matched_items": 2,
+        "predicted_items": 3,
+        "reflexive_entity_items": 0,
+        "rows": 1,
+        "target_items": 3,
+        "valid_predicted_items": 3
+      },
+      "document_id": "rail-incident-000",
+      "elapsed_ms": 16967,
+      "error": null,
+      "index": 0,
+      "parent_template": "incident-rail-style0",
+      "task": "metadata"
+    },
+    {
+      "arguments": {
+        "facts": [
+          {
+            "c": 0.99,
+            "h": false,
+            "o": "ex:rail-review-panel-000",
+            "p": "reportedTo",
+            "s": "ex:ari-alder-000"
+          },
+          {
+            "c": 0.99,
+            "h": false,
+            "o": true,
+            "p": "failed",
+            "s": "ex:rail-primary-system-000"
+          },
+          {
+            "c": 0.99,
+            "h": false,
+            "o": true,
+            "p": "remainedAvailable",
+            "s": "ex:rail-reserve-system-000"
+          },
+          {
+            "c": 0.99,
+            "h": false,
+            "o": "2032-01-01",
+            "p": "occurredOn",
+            "s": "ex:primary-event"
+          }
+        ],
+        "more_supported_facts": false
+      },
+      "counts": {
+        "calls_failed": 0,
+        "calls_ok": 1,
+        "duplicate_items": 0,
+        "invalid_predicted_items": 0,
+        "matched_items": 4,
+        "predicted_items": 4,
+        "reflexive_entity_items": 0,
+        "rows": 1,
+        "target_items": 4,
+        "valid_predicted_items": 4
+      },
+      "document_id": "rail-incident-000",
+      "elapsed_ms": 19304,
+      "error": null,
+      "index": 1,
+      "parent_template": "incident-rail-style0",
+      "task": "d0"
+    },
+    {
+      "arguments": {
+        "facts": [],
+        "more_supported_facts": false
+      },
+      "counts": {
+        "calls_failed": 0,
+        "calls_ok": 1,
+        "duplicate_items": 0,
+        "invalid_predicted_items": 0,
+        "matched_items": 0,
+        "predicted_items": 0,
+        "reflexive_entity_items": 0,
+        "rows": 1,
+        "target_items": 0,
+        "valid_predicted_items": 0
+      },
+      "document_id": "rail-incident-000",
+      "elapsed_ms": 6244,
+      "error": null,
+      "index": 2,
+      "parent_template": "incident-rail-style0",
+      "task": "residual_unit"
+    },
+    {
+      "arguments": {
+        "facts": [
+          {
+            "c": 0.99,
+            "h": false,
+            "o": "ex:hugo-lopez-000",
+            "p": "instructed",
+            "s": "ex:ari-alder-000"
+          },
+          {
+            "c": 0.99,
+            "h": false,
+            "o": "ex:task-rail-000",
+            "p": "instructedToPerform",
+            "s": "ex:hugo-lopez-000"
+          },
+          {
+            "c": 0.99,
+            "h": false,
+            "o": "ex:oona-wu-000",
+            "p": "requested",
+            "s": "ex:ari-alder-000"
+          },
+          {
+            "c": 0.99,
+            "h": false,
+            "o": "ex:status-display",
+            "p": "requestedToRecord",
+            "s": "ex:oona-wu-000"
+          }
+        ],
+        "more_supported_facts": false
+      },
+      "counts": {
+        "calls_failed": 0,
+        "calls_ok": 1,
+        "duplicate_items": 0,
+        "invalid_predicted_items": 0,
+        "matched_items": 4,
+        "predicted_items": 4,
+        "reflexive_entity_items": 0,
+        "rows": 1,
+        "target_items": 4,
+        "valid_predicted_items": 4
+      },
+      "document_id": "rail-incident-000",
+      "elapsed_ms": 20826,
+      "error": null,
+      "index": 3,
+      "parent_template": "incident-rail-style0",
+      "task": "d0"
+    },
+    {
+      "arguments": {
+        "facts": [],
+        "more_supported_facts": false
+      },
+      "counts": {
+        "calls_failed": 0,
+        "calls_ok": 1,
+        "duplicate_items": 0,
+        "invalid_predicted_items": 0,
+        "matched_items": 0,
+        "predicted_items": 0,
+        "reflexive_entity_items": 0,
+        "rows": 1,
+        "target_items": 0,
+        "valid_predicted_items": 0
+      },
+      "document_id": "rail-incident-000",
+      "elapsed_ms": 3512,
+      "error": null,
+      "index": 4,
+      "parent_template": "incident-rail-style0",
+      "task": "residual_unit"
+    },
+    {
+      "arguments": {
+        "facts": [
+          {
+            "c": 0.99,
+            "h": false,
+            "o": "ex:task-rail-000",
+            "p": "completed",
+            "s": "ex:hugo-lopez-000"
+          },
+          {
+            "c": 0.99,
+            "h": false,
+            "o": "07:00",
+            "p": "completedAt",
+            "s": "ex:task-rail-000"
+          },
+          {
+            "c": 0.99,
+            "h": false,
+            "o": "ex:omitted-step-rail-000",
+            "p": "didNotPerform",
+            "s": "ex:hugo-lopez-000"
+          },
+          {
+            "c": 0.99,
+            "h": false,
+            "o": "ex:effect-rail-000",
+            "p": "caused",
+            "s": "ex:omitted-step-rail-000"
+          }
+        ],
+        "more_supported_facts": false
+      },
+      "counts": {
+        "calls_failed": 0,
+        "calls_ok": 1,
+        "duplicate_items": 0,
+        "invalid_predicted_items": 0,
+        "matched_items": 4,
+        "predicted_items": 4,
+        "reflexive_entity_items": 0,
+        "rows": 1,
+        "target_items": 4,
+        "valid_predicted_items": 4
+      },
+      "document_id": "rail-incident-000",
+      "elapsed_ms": 18295,
+      "error": null,
+      "index": 5,
+      "parent_template": "incident-rail-style0",
+      "task": "d0"
+    },
+    {
+      "arguments": {
+        "facts": [],
+        "more_supported_facts": false
+      },
+      "counts": {
+        "calls_failed": 0,
+        "calls_ok": 1,
+        "duplicate_items": 0,
+        "invalid_predicted_items": 0,
+        "matched_items": 0,
+        "predicted_items": 0,
+        "reflexive_entity_items": 0,
+        "rows": 1,
+        "target_items": 0,
+        "valid_predicted_items": 0
+      },
+      "document_id": "rail-incident-000",
+      "elapsed_ms": 3695,
+      "error": null,
+      "index": 6,
+      "parent_template": "incident-rail-style0",
+      "task": "residual_unit"
+    },
+    {
+      "arguments": {
+        "facts": [
+          {
+            "c": 0.99,
+            "h": false,
+            "o": 40,
+            "p": "recordedValue",
+            "s": "ex:oona-wu-000"
+          },
+          {
+            "c": 0.99,
+            "h": false,
+            "o": "ex:ari-alder-000",
+            "p": "sentRecordTo",
+            "s": "ex:oona-wu-000"
+          },
+          {
+            "c": 0.99,
+            "h": false,
+            "o": "ex:deployment",
+            "p": "postponed",
+            "s": "ex:rail-review-panel-000"
+          },
+          {
+            "c": 0.99,
+            "h": false,
+            "o": "ex:independent-check-complete",
+            "p": "until",
+            "s": "ex:deployment-postponement"
+          }
+        ],
+        "more_supported_facts": false
+      },
+      "counts": {
+        "calls_failed": 0,
+        "calls_ok": 1,
+        "duplicate_items": 0,
+        "invalid_predicted_items": 0,
+        "matched_items": 4,
+        "predicted_items": 4,
+        "reflexive_entity_items": 0,
+        "rows": 1,
+        "target_items": 4,
+        "valid_predicted_items": 4
+      },
+      "document_id": "rail-incident-000",
+      "elapsed_ms": 19072,
+      "error": null,
+      "index": 7,
+      "parent_template": "incident-rail-style0",
+      "task": "d0"
+    },
+    {
+      "arguments": {
+        "facts": [],
+        "more_supported_facts": false
+      },
+      "counts": {
+        "calls_failed": 0,
+        "calls_ok": 1,
+        "duplicate_items": 0,
+        "invalid_predicted_items": 0,
+        "matched_items": 0,
+        "predicted_items": 0,
+        "reflexive_entity_items": 0,
+        "rows": 1,
+        "target_items": 0,
+        "valid_predicted_items": 0
+      },
+      "document_id": "rail-incident-000",
+      "elapsed_ms": 3674,
+      "error": null,
+      "index": 8,
+      "parent_template": "incident-rail-style0",
+      "task": "residual_unit"
+    },
+    {
+      "arguments": {
+        "facts": [
+          {
+            "c": 0.99,
+            "h": false,
+            "o": "ex:rail-deployment-postponement-000",
+            "p": "criticized",
+            "s": "ex:vera-hale-000"
+          },
+          {
+            "c": 0.99,
+            "h": false,
+            "o": "ex:rail-deployment-postponement-000",
+            "p": "defended",
+            "s": "ex:ari-alder-000"
+          },
+          {
+            "c": 0.99,
+            "h": false,
+            "o": "ex:rail-omitted-step-000",
+            "p": "attributed",
+            "s": "ex:rail-automated-alarm-000"
+          },
+          {
+            "c": 0.99,
+            "h": false,
+            "o": "ex:hugo-lopez-000",
+            "p": "didNotAttribute",
+            "s": "ex:rail-primary-failure-000"
+          }
+        ],
+        "more_supported_facts": false
+      },
+      "counts": {
+        "calls_failed": 0,
+        "calls_ok": 1,
+        "duplicate_items": 0,
+        "invalid_predicted_items": 0,
+        "matched_items": 0,
+        "predicted_items": 4,
+        "reflexive_entity_items": 0,
+        "rows": 1,
+        "target_items": 4,
+        "valid_predicted_items": 4
+      },
+      "document_id": "rail-incident-000",
+      "elapsed_ms": 21648,
+      "error": null,
+      "index": 9,
+      "parent_template": "incident-rail-style0",
+      "task": "d0"
+    },
+    {
+      "arguments": {
+        "facts": [],
+        "more_supported_facts": false
+      },
+      "counts": {
+        "calls_failed": 0,
+        "calls_ok": 1,
+        "duplicate_items": 0,
+        "invalid_predicted_items": 0,
+        "matched_items": 0,
+        "predicted_items": 0,
+        "reflexive_entity_items": 0,
+        "rows": 1,
+        "target_items": 0,
+        "valid_predicted_items": 0
+      },
+      "document_id": "rail-incident-000",
+      "elapsed_ms": 3542,
+      "error": null,
+      "index": 10,
+      "parent_template": "incident-rail-style0",
+      "task": "residual_unit"
+    },
+    {
+      "arguments": {
+        "facts": [
+          {
+            "c": 0.99,
+            "h": false,
+            "o": 40,
+            "p": "value",
+            "s": "ex:status-record"
+          },
+          {
+            "c": 0.99,
+            "h": false,
+            "o": "ex:oona-wu-000",
+            "p": "sentBy",
+            "s": "ex:status-record"
+          },
+          {
+            "c": 0.99,
+            "h": false,
+            "o": "ex:hugo-lopez-000",
+            "p": "notAttributedTo",
+            "s": "ex:original-failure"
+          }
+        ],
+        "more_supported_facts": false
+      },
+      "counts": {
+        "calls_failed": 0,
+        "calls_ok": 1,
+        "duplicate_items": 0,
+        "invalid_predicted_items": 0,
+        "matched_items": 3,
+        "predicted_items": 3,
+        "reflexive_entity_items": 0,
+        "rows": 1,
+        "target_items": 3,
+        "valid_predicted_items": 3
+      },
+      "document_id": "rail-incident-000",
+      "elapsed_ms": 15933,
+      "error": null,
+      "index": 11,
+      "parent_template": "incident-rail-style0",
+      "task": "residual"
+    },
+    {
+      "arguments": {
+        "facts": [],
+        "more_supported_facts": false
+      },
+      "counts": {
+        "calls_failed": 0,
+        "calls_ok": 1,
+        "duplicate_items": 0,
+        "invalid_predicted_items": 0,
+        "matched_items": 0,
+        "predicted_items": 0,
+        "reflexive_entity_items": 0,
+        "rows": 1,
+        "target_items": 0,
+        "valid_predicted_items": 0
+      },
+      "document_id": "rail-incident-000",
+      "elapsed_ms": 4883,
+      "error": null,
+      "index": 12,
+      "parent_template": "incident-rail-style0",
+      "task": "residual"
+    },
+    {
+      "arguments": {
+        "facts": [
+          {
+            "c": 0.92,
+            "h": true,
+            "o": "ex:ari-alder-000",
+            "p": "receivedReportFrom",
+            "s": "ex:rail-review-panel-000"
+          }
+        ],
+        "more_supported_facts": false
+      },
+      "counts": {
+        "calls_failed": 0,
+        "calls_ok": 1,
+        "duplicate_items": 0,
+        "invalid_predicted_items": 0,
+        "matched_items": 1,
+        "predicted_items": 1,
+        "reflexive_entity_items": 0,
+        "rows": 1,
+        "target_items": 1,
+        "valid_predicted_items": 1
+      },
+      "document_id": "rail-incident-000",
+      "elapsed_ms": 7981,
+      "error": null,
+      "index": 13,
+      "parent_template": "incident-rail-style0",
+      "task": "d1_unit"
+    },
+    {
+      "arguments": {
+        "facts": [],
+        "more_supported_facts": false
+      },
+      "counts": {
+        "calls_failed": 0,
+        "calls_ok": 1,
+        "duplicate_items": 0,
+        "invalid_predicted_items": 0,
+        "matched_items": 0,
+        "predicted_items": 0,
+        "reflexive_entity_items": 0,
+        "rows": 1,
+        "target_items": 0,
+        "valid_predicted_items": 0
+      },
+      "document_id": "rail-incident-000",
+      "elapsed_ms": 3931,
+      "error": null,
+      "index": 14,
+      "parent_template": "incident-rail-style0",
+      "task": "d1_unit"
+    },
+    {
+      "arguments": {
+        "facts": [
+          {
+            "c": 0.92,
+            "h": true,
+            "o": "ex:ari-alder-000",
+            "p": "receivedInstructionFrom",
+            "s": "ex:hugo-lopez-000"
+          },
+          {
+            "c": 0.92,
+            "h": true,
+            "o": "ex:ari-alder-000",
+            "p": "receivedRequestFrom",
+            "s": "ex:oona-wu-000"
+          },
+          {
+            "c": 0.92,
+            "h": true,
+            "o": "ex:oona-wu-000",
+            "p": "receivedRecordFrom",
+            "s": "ex:ari-alder-000"
+          }
+        ],
+        "more_supported_facts": false
+      },
+      "counts": {
+        "calls_failed": 0,
+        "calls_ok": 1,
+        "duplicate_items": 0,
+        "invalid_predicted_items": 0,
+        "matched_items": 3,
+        "predicted_items": 3,
+        "reflexive_entity_items": 0,
+        "rows": 1,
+        "target_items": 3,
+        "valid_predicted_items": 3
+      },
+      "document_id": "rail-incident-000",
+      "elapsed_ms": 19858,
+      "error": null,
+      "index": 15,
+      "parent_template": "incident-rail-style0",
+      "task": "d1_unit"
+    },
+    {
+      "arguments": {
+        "facts": [],
+        "more_supported_facts": false
+      },
+      "counts": {
+        "calls_failed": 0,
+        "calls_ok": 1,
+        "duplicate_items": 0,
+        "invalid_predicted_items": 0,
+        "matched_items": 0,
+        "predicted_items": 0,
+        "reflexive_entity_items": 0,
+        "rows": 1,
+        "target_items": 0,
+        "valid_predicted_items": 0
+      },
+      "document_id": "rail-incident-000",
+      "elapsed_ms": 4232,
+      "error": null,
+      "index": 16,
+      "parent_template": "incident-rail-style0",
+      "task": "d1_unit"
+    },
+    {
+      "arguments": {
+        "facts": [
+          {
+            "c": 0.92,
+            "h": true,
+            "o": "ex:omitted-step-rail-000",
+            "p": "causedBy",
+            "s": "ex:effect-rail-000"
+          }
+        ],
+        "more_supported_facts": false
+      },
+      "counts": {
+        "calls_failed": 0,
+        "calls_ok": 1,
+        "duplicate_items": 0,
+        "invalid_predicted_items": 0,
+        "matched_items": 1,
+        "predicted_items": 1,
+        "reflexive_entity_items": 0,
+        "rows": 1,
+        "target_items": 1,
+        "valid_predicted_items": 1
+      },
+      "document_id": "rail-incident-000",
+      "elapsed_ms": 9552,
+      "error": null,
+      "index": 17,
+      "parent_template": "incident-rail-style0",
+      "task": "d1_unit"
+    },
+    {
+      "arguments": {
+        "facts": [],
+        "more_supported_facts": false
+      },
+      "counts": {
+        "calls_failed": 0,
+        "calls_ok": 1,
+        "duplicate_items": 0,
+        "invalid_predicted_items": 0,
+        "matched_items": 0,
+        "predicted_items": 0,
+        "reflexive_entity_items": 0,
+        "rows": 1,
+        "target_items": 0,
+        "valid_predicted_items": 0
+      },
+      "document_id": "rail-incident-000",
+      "elapsed_ms": 3905,
+      "error": null,
+      "index": 18,
+      "parent_template": "incident-rail-style0",
+      "task": "d1_unit"
+    },
+    {
+      "arguments": {
+        "facts": [],
+        "more_supported_facts": false
+      },
+      "counts": {
+        "calls_failed": 0,
+        "calls_ok": 1,
+        "duplicate_items": 0,
+        "invalid_predicted_items": 0,
+        "matched_items": 0,
+        "predicted_items": 0,
+        "reflexive_entity_items": 0,
+        "rows": 1,
+        "target_items": 0,
+        "valid_predicted_items": 0
+      },
+      "document_id": "rail-incident-000",
+      "elapsed_ms": 3820,
+      "error": null,
+      "index": 19,
+      "parent_template": "incident-rail-style0",
+      "task": "d1_unit"
+    },
+    {
+      "arguments": {
+        "facts": [],
+        "more_supported_facts": false
+      },
+      "counts": {
+        "calls_failed": 0,
+        "calls_ok": 1,
+        "duplicate_items": 0,
+        "invalid_predicted_items": 0,
+        "matched_items": 0,
+        "predicted_items": 0,
+        "reflexive_entity_items": 0,
+        "rows": 1,
+        "target_items": 0,
+        "valid_predicted_items": 0
+      },
+      "document_id": "rail-incident-000",
+      "elapsed_ms": 5779,
+      "error": null,
+      "index": 20,
+      "parent_template": "incident-rail-style0",
+      "task": "d1_unit"
+    },
+    {
+      "arguments": {
+        "facts": [],
+        "more_supported_facts": false
+      },
+      "counts": {
+        "calls_failed": 0,
+        "calls_ok": 1,
+        "duplicate_items": 0,
+        "invalid_predicted_items": 0,
+        "matched_items": 0,
+        "predicted_items": 0,
+        "reflexive_entity_items": 0,
+        "rows": 1,
+        "target_items": 0,
+        "valid_predicted_items": 0
+      },
+      "document_id": "rail-incident-000",
+      "elapsed_ms": 5970,
+      "error": null,
+      "index": 21,
+      "parent_template": "incident-rail-style0",
+      "task": "d1"
+    },
+    {
+      "arguments": {
+        "facts": [
+          {
+            "c": 0.92,
+            "h": true,
+            "o": "ex:postponement-protected-staff",
+            "p": "contrastsWith",
+            "s": "ex:postponement-wasted-money"
+          },
+          {
+            "c": 0.92,
+            "h": true,
+            "o": "ex:effect-rail-000",
+            "p": "explains",
+            "s": "ex:omitted-step-rail-000"
+          },
+          {
+            "c": 0.92,
+            "h": true,
+            "o": "ex:deployment",
+            "p": "isPreconditionFor",
+            "s": "ex:independent-check"
+          }
+        ],
+        "more_supported_facts": false
+      },
+      "counts": {
+        "calls_failed": 0,
+        "calls_ok": 1,
+        "duplicate_items": 0,
+        "invalid_predicted_items": 0,
+        "matched_items": 3,
+        "predicted_items": 3,
+        "reflexive_entity_items": 0,
+        "rows": 1,
+        "target_items": 3,
+        "valid_predicted_items": 3
+      },
+      "document_id": "rail-incident-000",
+      "elapsed_ms": 17798,
+      "error": null,
+      "index": 22,
+      "parent_template": "incident-rail-style0",
+      "task": "d2"
+    },
+    {
+      "arguments": {
+        "facts": [],
+        "more_supported_facts": false
+      },
+      "counts": {
+        "calls_failed": 0,
+        "calls_ok": 1,
+        "duplicate_items": 0,
+        "invalid_predicted_items": 0,
+        "matched_items": 0,
+        "predicted_items": 0,
+        "reflexive_entity_items": 0,
+        "rows": 1,
+        "target_items": 0,
+        "valid_predicted_items": 0
+      },
+      "document_id": "rail-incident-000",
+      "elapsed_ms": 4707,
+      "error": null,
+      "index": 23,
+      "parent_template": "incident-rail-style0",
+      "task": "d2"
+    },
+    {
+      "arguments": {
+        "facts": [
+          {
+            "c": 0.92,
+            "h": true,
+            "o": "ex:Person",
+            "p": "rdf:type",
+            "s": "ex:ari-alder-000"
+          },
+          {
+            "c": 0.99,
+            "h": false,
+            "o": "Ari Alder-000",
+            "p": "rdfs:label",
+            "s": "ex:ari-alder-000"
+          },
+          {
+            "c": 0.92,
+            "h": true,
+            "o": "ex:Person",
+            "p": "rdf:type",
+            "s": "ex:hugo-lopez-000"
+          },
+          {
+            "c": 0.99,
+            "h": false,
+            "o": "Hugo Lopez-000",
+            "p": "rdfs:label",
+            "s": "ex:hugo-lopez-000"
+          },
+          {
+            "c": 0.92,
+            "h": true,
+            "o": "ex:Person",
+            "p": "rdf:type",
+            "s": "ex:oona-wu-000"
+          },
+          {
+            "c": 0.99,
+            "h": false,
+            "o": "Oona Wu-000",
+            "p": "rdfs:label",
+            "s": "ex:oona-wu-000"
+          },
+          {
+            "c": 0.92,
+            "h": true,
+            "o": "ex:Event",
+            "p": "rdf:type",
+            "s": "ex:primary-event"
+          },
+          {
+            "c": 0.92,
+            "h": true,
+            "o": "primary event",
+            "p": "rdfs:label",
+            "s": "ex:primary-event"
+          },
+          {
+            "c": 0.92,
+            "h": true,
+            "o": "ex:SyntheticTrainingDocument",
+            "p": "rdf:type",
+            "s": "ex:rail-operations-review-000"
+          },
+          {
+            "c": 0.99,
+            "h": false,
+            "o": "Rail Operations Review 000",
+            "p": "rdfs:label",
+            "s": "ex:rail-operations-review-000"
+          },
+          {
+            "c": 0.92,
+            "h": true,
+            "o": "ex:Organization",
+            "p": "rdf:type",
+            "s": "ex:rail-review-panel-000"
+          },
+          {
+            "c": 0.99,
+            "h": false,
+            "o": "Rail Review Panel 000",
+            "p": "rdfs:label",
+            "s": "ex:rail-review-panel-000"
+          }
+        ],
+        "more_supported_facts": false
+      },
+      "counts": {
+        "calls_failed": 0,
+        "calls_ok": 1,
+        "duplicate_items": 0,
+        "invalid_predicted_items": 0,
+        "matched_items": 12,
+        "predicted_items": 12,
+        "reflexive_entity_items": 0,
+        "rows": 1,
+        "target_items": 12,
+        "valid_predicted_items": 12
+      },
+      "document_id": "rail-incident-000",
+      "elapsed_ms": 37290,
+      "error": null,
+      "index": 24,
+      "parent_template": "incident-rail-style0",
+      "task": "structural_batch"
+    },
+    {
+      "arguments": {
+        "facts": [],
+        "more_supported_facts": false
+      },
+      "counts": {
+        "calls_failed": 0,
+        "calls_ok": 1,
+        "duplicate_items": 0,
+        "invalid_predicted_items": 0,
+        "matched_items": 0,
+        "predicted_items": 0,
+        "reflexive_entity_items": 0,
+        "rows": 1,
+        "target_items": 0,
+        "valid_predicted_items": 0
+      },
+      "document_id": "rail-incident-000",
+      "elapsed_ms": 4333,
+      "error": null,
+      "index": 25,
+      "parent_template": "incident-rail-style0",
+      "task": "structural_batch"
+    },
+    {
+      "arguments": {
+        "facts": [
+          {
+            "c": 0.99,
+            "h": false,
+            "o": "ex:dara-xu-001",
+            "p": "authoredBy",
+            "s": "ex:university-testimony-digest-001"
+          },
+          {
+            "c": 0.99,
+            "h": false,
+            "o": "2030",
+            "p": "dated",
+            "s": "ex:university-testimony-digest-001"
+          },
+          {
+            "c": 0.99,
+            "h": false,
+            "o": "ex:committee-hearing",
+            "p": "describes",
+            "s": "ex:university-testimony-digest-001"
+          }
+        ],
+        "more_supported_facts": false
+      },
+      "counts": {
+        "calls_failed": 0,
+        "calls_ok": 1,
+        "duplicate_items": 0,
+        "invalid_predicted_items": 0,
+        "matched_items": 3,
+        "predicted_items": 3,
+        "reflexive_entity_items": 0,
+        "rows": 1,
+        "target_items": 3,
+        "valid_predicted_items": 3
+      },
+      "document_id": "university-quote-001",
+      "elapsed_ms": 16390,
+      "error": null,
+      "index": 26,
+      "parent_template": "quote-university-style0",
+      "task": "metadata"
+    },
+    {
+      "arguments": {
+        "facts": [
+          {
+            "c": 0.99,
+            "h": false,
+            "o": "ex:dara-xu-001",
+            "p": "authoredBy",
+            "s": "ex:university-testimony-digest-001"
+          },
+          {
+            "c": 0.99,
+            "h": false,
+            "o": "ex:committee-hearing",
+            "p": "describes",
+            "s": "ex:university-testimony-digest-001"
+          },
+          {
+            "c": 0.99,
+            "h": false,
+            "o": "ex:university-committee-001",
+            "p": "heldBy",
+            "s": "ex:committee-hearing"
+          },
+          {
+            "c": 0.99,
+            "h": false,
+            "o": "2029-02-02",
+            "p": "heldOn",
+            "s": "ex:committee-hearing"
+          }
+        ],
+        "more_supported_facts": false
+      },
+      "counts": {
+        "calls_failed": 0,
+        "calls_ok": 1,
+        "duplicate_items": 0,
+        "invalid_predicted_items": 0,
+        "matched_items": 4,
+        "predicted_items": 4,
+        "reflexive_entity_items": 0,
+        "rows": 1,
+        "target_items": 4,
+        "valid_predicted_items": 4
+      },
+      "document_id": "university-quote-001",
+      "elapsed_ms": 19404,
+      "error": null,
+      "index": 27,
+      "parent_template": "quote-university-style0",
+      "task": "d0"
+    },
+    {
+      "arguments": {
+        "facts": [],
+        "more_supported_facts": false
+      },
+      "counts": {
+        "calls_failed": 0,
+        "calls_ok": 1,
+        "duplicate_items": 0,
+        "invalid_predicted_items": 0,
+        "matched_items": 0,
+        "predicted_items": 0,
+        "reflexive_entity_items": 0,
+        "rows": 1,
+        "target_items": 0,
+        "valid_predicted_items": 0
+      },
+      "document_id": "university-quote-001",
+      "elapsed_ms": 4078,
+      "error": null,
+      "index": 28,
+      "parent_template": "quote-university-style0",
+      "task": "residual_unit"
+    },
+    {
+      "arguments": {
+        "facts": [
+          {
+            "c": 0.99,
+            "h": false,
+            "o": "ex:speaker-statement",
+            "p": "reports",
+            "s": "ex:dara-xu-001"
+          },
+          {
+            "c": 0.99,
+            "h": false,
+            "o": "ex:rhea-tran-001",
+            "p": "told",
+            "s": "ex:kira-ito-001"
+          },
+          {
+            "c": 0.99,
+            "h": false,
+            "o": "ex:external-review",
+            "p": "approved",
+            "s": "ex:kira-ito-001"
+          },
+          {
+            "c": 0.99,
+            "h": false,
+            "o": "ex:university-proposal-001",
+            "p": "didNotApprove",
+            "s": "ex:kira-ito-001"
+          }
+        ],
+        "more_supported_facts": false
+      },
+      "counts": {
+        "calls_failed": 0,
+        "calls_ok": 1,
+        "duplicate_items": 0,
+        "invalid_predicted_items": 0,
+        "matched_items": 4,
+        "predicted_items": 4,
+        "reflexive_entity_items": 0,
+        "rows": 1,
+        "target_items": 4,
+        "valid_predicted_items": 4
+      },
+      "document_id": "university-quote-001",
+      "elapsed_ms": 14177,
+      "error": null,
+      "index": 29,
+      "parent_template": "quote-university-style0",
+      "task": "d0"
+    }
+  ]
+}

examples/structured_client.py

@@ -0,0 +1,65 @@
+#!/usr/bin/env python3
+"""Forced-tool example for a running Donto-Qwen vLLM endpoint."""
+
+from typing import Union
+
+from openai import OpenAI
+from pydantic import BaseModel, ConfigDict, Field
+
+
+class Fact(BaseModel):
+    model_config = ConfigDict(extra="forbid")
+
+    s: str
+    p: str
+    o: Union[str, int, float, bool]
+    c: float = Field(ge=0.0, le=1.0)
+    h: bool
+
+
+class Submission(BaseModel):
+    model_config = ConfigDict(extra="forbid")
+
+    facts: list[Fact] = Field(max_length=64)
+    more_supported_facts: bool
+
+
+TOOLS = [
+    {
+        "type": "function",
+        "function": {
+            "name": "submit_facts",
+            "description": "Submit validated Donto extraction facts.",
+            "parameters": Submission.model_json_schema(),
+        },
+    }
+]
+
+SOURCE = """Ari reported that the primary system failed, while the reserve
+system remained available. The panel postponed deployment until an independent
+check was complete. Vera said the postponement wasted money, but Ari argued that
+it protected staff."""
+
+PROMPT = f"""Extract every atomic proposition directly supported by SOURCE.
+Preserve attribution, contrast, modality, and negation. Use stable ex:kebab-case
+entities, concise camelCase predicates, h:false for direct facts, and no prose.
+
+SOURCE:
+{SOURCE}
+"""
+
+client = OpenAI(base_url="http://127.0.0.1:8000/v1", api_key="unused")
+response = client.chat.completions.create(
+    model="donto-qwen",
+    messages=[{"role": "user", "content": PROMPT}],
+    tools=TOOLS,
+    tool_choice={"type": "function", "function": {"name": "submit_facts"}},
+    temperature=0,
+    max_tokens=1024,
+    extra_body={"chat_template_kwargs": {"enable_thinking": False, "preserve_thinking": False}},
+)
+
+call = response.choices[0].message.tool_calls[0]
+submission = Submission.model_validate_json(call.function.arguments)
+print(submission.model_dump_json(indent=2))
+

examples/vllm_serve.sh

@@ -0,0 +1,24 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+ADAPTER_DIR=${ADAPTER_DIR:-./donto-qwen/runtime/vllm-conditional}
+
+vllm serve Qwen/Qwen3.8-27B \
+  --served-model-name qwen3.8-27b \
+  --language-model-only \
+  --quantization bitsandbytes \
+  --load-format bitsandbytes \
+  --dtype bfloat16 \
+  --max-model-len 4096 \
+  --max-num-seqs 4 \
+  --max-num-batched-tokens 2048 \
+  --enable-prefix-caching \
+  --generation-config vllm \
+  --reasoning-parser qwen3 \
+  --enable-auto-tool-choice \
+  --tool-call-parser qwen3_coder \
+  --enable-lora \
+  --lora-modules "donto-qwen=${ADAPTER_DIR}" \
+  --max-loras 1 \
+  --max-lora-rank 16
+

requirements.txt

@@ -0,0 +1,11 @@
+accelerate==1.14.0
+bitsandbytes==0.50.1
+huggingface-hub>=1.28.0
+peft==0.20.0
+pydantic>=2.0
+openai>=1.0
+safetensors>=0.4
+torch>=2.8.0
+transformers==5.15.0
+vllm==0.27.1
+

runtime/vllm-conditional/REMAPPING-MANIFEST.json

@@ -0,0 +1,30 @@
+{
+  "blind_holdout_opened": false,
+  "created_at": "2026-08-25T09:14:27.332868+00:00",
+  "original_target_modules": ".*(?:language_model\\.)?layers\\.\\d+\\.(?:self_attn\\.(?:q_proj|k_proj|v_proj|o_proj)|linear_attn\\.(?:in_proj_qkv|in_proj_z|in_proj_a|in_proj_b|out_proj)|mlp\\.(?:gate_proj|up_proj|down_proj))$",
+  "output": "/workspace/donto-qwen80/pilot-live-v15-exhaustive-unit-r16-lr2e5-seed38-v1/vllm-conditional",
+  "output_adapter_sha256": "28df663ffa70bf541cafff2c9bb60df11e4fe000a2ea9124b346a9cc3eb20d8b",
+  "output_config_sha256": "1f8abcefc88d6db74239c889fd2246c94b45da3b58c9d68ba8360692b39f06a2",
+  "schema": "donto-qwen38-vllm-conditional-adapter-remap-v1",
+  "source": "/workspace/donto-qwen80/pilot-live-v15-exhaustive-unit-r16-lr2e5-seed38-v1/final",
+  "source_adapter_sha256": "40d11bcad20061f8291741c4dd8a3701405902f336c27b60592d62d6b2f7a974",
+  "source_config_sha256": "a928e2f4694e2145d7e3c6c7f892bcdc9591a45baa6fa97e380ccb01d495b875",
+  "source_key_prefix": "base_model.model.model.layers.",
+  "target_key_prefix": "base_model.model.model.language_model.layers.",
+  "tensor_count": 992,
+  "tensor_values_changed": false,
+  "vllm_target_modules": [
+    "q_proj",
+    "k_proj",
+    "v_proj",
+    "o_proj",
+    "in_proj_qkv",
+    "in_proj_z",
+    "in_proj_a",
+    "in_proj_b",
+    "out_proj",
+    "gate_proj",
+    "up_proj",
+    "down_proj"
+  ]
+}

runtime/vllm-conditional/adapter_config.json

@@ -0,0 +1,55 @@
+{
+  "alora_invocation_tokens": null,
+  "alpha_pattern": {},
+  "arrow_config": null,
+  "auto_mapping": null,
+  "base_model_name_or_path": "Qwen/Qwen3.8-27B",
+  "bias": "none",
+  "corda_config": null,
+  "ensure_weight_tying": false,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 32,
+  "lora_bias": false,
+  "lora_dropout": 0.05,
+  "lora_ga_config": null,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "monteclora_config": null,
+  "peft_type": "LORA",
+  "peft_version": "0.20.0",
+  "qalora_group_size": 16,
+  "r": 16,
+  "rank_pattern": {},
+  "revision": "1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0",
+  "target_modules": [
+    "q_proj",
+    "k_proj",
+    "v_proj",
+    "o_proj",
+    "in_proj_qkv",
+    "in_proj_z",
+    "in_proj_a",
+    "in_proj_b",
+    "out_proj",
+    "gate_proj",
+    "up_proj",
+    "down_proj"
+  ],
+  "target_parameters": null,
+  "task_type": "CAUSAL_LM",
+  "trainable_token_indices": null,
+  "use_bdlora": null,
+  "use_dora": false,
+  "use_qalora": false,
+  "use_rslora": false,
+  "velora_config": null
+}

runtime/vllm-conditional/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:28df663ffa70bf541cafff2c9bb60df11e4fe000a2ea9124b346a9cc3eb20d8b
+size 467062560

runtime/vllm-conditional/chat_template.jinja

@@ -0,0 +1,170 @@
+{%- set image_count = namespace(value=0) %}
+{%- set video_count = namespace(value=0) %}
+{%- macro render_content(content, do_vision_count, is_system_content=false) %}
+    {%- if content is string %}
+        {{- content }}
+    {%- elif content is iterable and content is not mapping %}
+        {%- for item in content %}
+            {%- if 'image' in item or 'image_url' in item or item.type == 'image' %}
+                {%- if is_system_content %}
+                    {{- raise_exception('System message cannot contain images.') }}
+                {%- endif %}
+                {%- if do_vision_count %}
+                    {%- set image_count.value = image_count.value + 1 %}
+                {%- endif %}
+                {%- if add_vision_id %}
+                    {{- 'Picture ' ~ image_count.value ~ ': ' }}
+                {%- endif %}
+                {{- '<|vision_start|><|image_pad|><|vision_end|>' }}
+            {%- elif 'video' in item or item.type == 'video' %}
+                {%- if is_system_content %}
+                    {{- raise_exception('System message cannot contain videos.') }}
+                {%- endif %}
+                {%- if do_vision_count %}
+                    {%- set video_count.value = video_count.value + 1 %}
+                {%- endif %}
+                {%- if add_vision_id %}
+                    {{- 'Video ' ~ video_count.value ~ ': ' }}
+                {%- endif %}
+                {{- '<|vision_start|><|video_pad|><|vision_end|>' }}
+            {%- elif 'text' in item %}
+                {{- item.text }}
+            {%- else %}
+                {{- raise_exception('Unexpected item type in content.') }}
+            {%- endif %}
+        {%- endfor %}
+    {%- elif content is none or content is undefined %}
+        {{- '' }}
+    {%- else %}
+        {{- raise_exception('Unexpected content type.') }}
+    {%- endif %}
+{%- endmacro %}
+{%- if not messages %}
+    {{- raise_exception('No messages provided.') }}
+{%- endif %}
+{%- set reasoning_instructions = '' %}
+{%- if enable_thinking is undefined or enable_thinking is true %}
+    {%- set resolved_reasoning_effort = reasoning_effort|default('xhigh') %}
+    {%- if resolved_reasoning_effort not in ('xhigh', 'medium', 'low') %}
+        {{- raise_exception('Unexpected reasoning effort ' ~ reasoning_effort ~ '. Supported types are xhigh (default), medium, and low.') }}
+    {%- endif %}
+    {%- if resolved_reasoning_effort == 'xhigh' %}
+        {%- set reasoning_instructions = 'Reasoning effort is set to xhigh. Please think carefully through the task, validate key assumptions, consider plausible alternatives, and prioritize correctness, consistency, and clarity in the final answer.' %}
+    {%- elif resolved_reasoning_effort == 'low' %}
+        {%- set reasoning_instructions = 'Reasoning effort is set to low. Keep your thinking brief and focused, moving directly to the conclusion without unnecessary elaboration.' %}
+    {%- endif %}
+{%- endif %}
+{%- if tools and tools is iterable and tools is not mapping %}
+    {{- '<|im_start|>system\n' }}
+    {%- if reasoning_instructions %}
+        {{- reasoning_instructions + '\n\n' }}
+    {%- endif %}
+    {{- "# Tools\n\nYou have access to the following functions:\n\n<tools>" }}
+    {%- for tool in tools %}
+        {{- "\n" }}
+        {{- tool | tojson }}
+    {%- endfor %}
+    {{- "\n</tools>" }}
+    {{- '\n\nIf you choose to call a function ONLY reply in the following format with NO suffix:\n\n<tool_call>\n<function=example_function_name>\n<parameter=example_parameter_1>\nvalue_1\n</parameter>\n<parameter=example_parameter_2>\nThis is the value for the second parameter\nthat can span\nmultiple lines\n</parameter>\n</function>\n</tool_call>\n\n<IMPORTANT>\nReminder:\n- Function calls MUST follow the specified format: an inner <function=...></function> block must be nested within <tool_call></tool_call> XML tags\n- Required parameters MUST be specified\n- You may provide optional reasoning for your function call in natural language BEFORE the function call, but NOT after\n- If there is no function call available, answer the question like normal with your current knowledge and do not tell the user about function calls\n</IMPORTANT>' }}
+    {%- if messages[0].role == 'system' %}
+        {%- set content = render_content(messages[0].content, false, true)|trim %}
+        {%- if content %}
+            {{- '\n\n' + content }}
+        {%- endif %}
+    {%- endif %}
+    {{- '<|im_end|>\n' }}
+{%- else %}
+    {%- if messages[0].role == 'system' %}
+        {%- set content = render_content(messages[0].content, false, true)|trim %}
+        {%- if content %}
+            {{- '<|im_start|>system\n' + (reasoning_instructions + '\n\n' if reasoning_instructions else '')  + content + '<|im_end|>\n' }}
+        {%- elif reasoning_instructions %}
+            {{- '<|im_start|>system\n' + reasoning_instructions + '<|im_end|>\n' }}
+        {%- endif %}
+    {%- elif reasoning_instructions %}
+        {{- '<|im_start|>system\n' + reasoning_instructions + '<|im_end|>\n' }}
+    {%- endif %}
+{%- endif %}
+{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}
+{%- for message in messages[::-1] %}
+    {%- set index = (messages|length - 1) - loop.index0 %}
+    {%- if ns.multi_step_tool and message.role == "user" %}
+        {%- set content = render_content(message.content, false)|trim %}
+        {%- if not(content.startswith('<tool_response>') and content.endswith('</tool_response>')) %}
+            {%- set ns.multi_step_tool = false %}
+            {%- set ns.last_query_index = index %}
+        {%- endif %}
+    {%- endif %}
+{%- endfor %}
+{%- if ns.multi_step_tool %}
+    {{- raise_exception('No user query found in messages.') }}
+{%- endif %}
+{%- for message in messages %}
+    {%- set content = render_content(message.content, true)|trim %}
+    {%- if message.role == "system" %}
+        {%- if not loop.first %}
+            {{- raise_exception('System message must be at the beginning.') }}
+        {%- endif %}
+    {%- elif message.role == "user" %}
+        {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {%- set reasoning_content = '' %}
+        {%- if message.reasoning_content is string %}
+            {%- set reasoning_content = message.reasoning_content %}
+        {%- endif %}
+        {%- set reasoning_content = reasoning_content|trim %}
+        {%- if preserve_thinking is undefined or preserve_thinking is true or loop.index0 > ns.last_query_index %}
+            {{- '<|im_start|>' + message.role + '\n<think>\n' + reasoning_content + '\n</think>\n\n' + content }}
+        {%- else %}
+            {{- '<|im_start|>' + message.role + '\n' + content }}
+        {%- endif %}
+        {%- if message.tool_calls and message.tool_calls is iterable and message.tool_calls is not mapping %}
+            {%- for tool_call in message.tool_calls %}
+                {%- if tool_call.function is defined %}
+                    {%- set tool_call = tool_call.function %}
+                {%- endif %}
+                {%- if loop.first %}
+                    {%- if content|trim %}
+                        {{- '\n\n<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                    {%- else %}
+                        {{- '<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                    {%- endif %}
+                {%- else %}
+                    {{- '\n<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                {%- endif %}
+                {%- if tool_call.arguments is defined and tool_call.arguments != '' %}
+                    {%- for args_name, args_value in tool_call.arguments|items %}
+                        {{- '<parameter=' + args_name + '>\n' }}
+                        {%- set args_value = args_value | string if args_value is string else args_value | tojson | safe %}
+                        {{- args_value }}
+                        {{- '\n</parameter>\n' }}
+                    {%- endfor %}
+                {%- endif %}
+                {{- '</function>\n</tool_call>' }}
+            {%- endfor %}
+        {%- endif %}
+        {{- '<|im_end|>\n' }}
+    {%- elif message.role == "tool" %}
+        {%- if loop.previtem and loop.previtem.role != "tool" %}
+            {{- '<|im_start|>user' }}
+        {%- endif %}
+        {{- '\n<tool_response>\n' }}
+        {{- content }}
+        {{- '\n</tool_response>' }}
+        {%- if not loop.last and loop.nextitem.role != "tool" %}
+            {{- '<|im_end|>\n' }}
+        {%- elif loop.last %}
+            {{- '<|im_end|>\n' }}
+        {%- endif %}
+    {%- else %}
+        {{- raise_exception('Unexpected message role.') }}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|im_start|>assistant\n' }}
+    {%- if enable_thinking is defined and enable_thinking is false %}
+        {{- '<think>\n\n</think>\n\n' }}
+    {%- else %}
+        {{- '<think>\n' }}
+    {%- endif %}
+{%- endif %}

schemas/tools.json

@@ -0,0 +1,72 @@
+[
+  {
+    "function": {
+      "description": "Submit validated donto extraction facts for the source document.",
+      "name": "submit_facts",
+      "parameters": {
+        "additionalProperties": false,
+        "properties": {
+          "facts": {
+            "items": {
+              "additionalProperties": false,
+              "properties": {
+                "c": {
+                  "description": "Confidence from 0 to 1.",
+                  "maximum": 1,
+                  "minimum": 0,
+                  "type": "number"
+                },
+                "h": {
+                  "description": "True when inferred/unanchorable/hypothesis.",
+                  "type": "boolean"
+                },
+                "o": {
+                  "anyOf": [
+                    {
+                      "type": "string"
+                    },
+                    {
+                      "type": "number"
+                    },
+                    {
+                      "type": "boolean"
+                    }
+                  ],
+                  "description": "Object: entity IRI string, natural literal string, number, or boolean."
+                },
+                "p": {
+                  "description": "Precise camelCase predicate, or canonical rdf:type/rdfs:label/skos:broader/sameAs.",
+                  "type": "string"
+                },
+                "s": {
+                  "description": "Subject IRI, usually ex:/ctx:/http(s).",
+                  "type": "string"
+                }
+              },
+              "required": [
+                "s",
+                "p",
+                "o",
+                "c",
+                "h"
+              ],
+              "type": "object"
+            },
+            "maxItems": 64,
+            "type": "array"
+          },
+          "more_supported_facts": {
+            "description": "True only when another page of genuinely new, source-supported facts remains after this page.",
+            "type": "boolean"
+          }
+        },
+        "required": [
+          "facts",
+          "more_supported_facts"
+        ],
+        "type": "object"
+      }
+    },
+    "type": "function"
+  }
+]

tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:06b9509352d2af50381ab2247e083b80d32d5c0aba91c272ca9ff729b6a0e523
+size 19989325

tokenizer_config.json

@@ -0,0 +1,32 @@
+{
+  "add_prefix_space": false,
+  "audio_bos_token": "<|audio_start|>",
+  "audio_eos_token": "<|audio_end|>",
+  "audio_token": "<|audio_pad|>",
+  "backend": "tokenizers",
+  "bos_token": null,
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "image_token": "<|image_pad|>",
+  "is_local": true,
+  "local_files_only": false,
+  "model_max_length": 262144,
+  "model_specific_special_tokens": {
+    "audio_bos_token": "<|audio_start|>",
+    "audio_eos_token": "<|audio_end|>",
+    "audio_token": "<|audio_pad|>",
+    "image_token": "<|image_pad|>",
+    "video_token": "<|video_pad|>",
+    "vision_bos_token": "<|vision_start|>",
+    "vision_eos_token": "<|vision_end|>"
+  },
+  "pad_token": "<|endoftext|>",
+  "pretokenize_regex": "(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\\r\\n\\p{L}\\p{N}]?[\\p{L}\\p{M}]+|\\p{N}| ?[^\\s\\p{L}\\p{M}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null,
+  "video_token": "<|video_pad|>",
+  "vision_bos_token": "<|vision_start|>",
+  "vision_eos_token": "<|vision_end|>"
+}

training/checkpoints/checkpoint-100/CHECKSUMS.sha256

@@ -0,0 +1,8 @@
+df9859f741f75a96a958b358a1c08817e5fa6ff4a8fe8eed85a01add501ae7f1  README.md
+a928e2f4694e2145d7e3c6c7f892bcdc9591a45baa6fa97e380ccb01d495b875  adapter_config.json
+ac0fc8a29bdf8bd0de70dd275333cd6cd82e328a161ea86ac44e5e956bb85e67  adapter_model.safetensors
+dc9b5826d72f74051d11212ee12965cbeb2654b4d262f8f26defca7c8c614dfc  optimizer.pt
+4f4fb89bca81c597f60365bb7d285322d907cd99644e7fdd098368a51d8c8a57  rng_state.pth
+61f4fb0f34daf13abf304a22501429019f5b65ec145513f2a2ded5685f8e4142  scheduler.pt
+3257b63824c8c43da377913156b18400525c69caea5bca71c671d6eb49b5a75c  trainer_state.json
+175111135c9d40b63054d1ef321345f80de8c4d21717241624ae93bced357d6c  training_args.bin

training/checkpoints/checkpoint-100/adapter_config.json

@@ -0,0 +1,42 @@
+{
+  "alora_invocation_tokens": null,
+  "alpha_pattern": {},
+  "arrow_config": null,
+  "auto_mapping": null,
+  "base_model_name_or_path": "Qwen/Qwen3.8-27B",
+  "bias": "none",
+  "corda_config": null,
+  "ensure_weight_tying": false,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 32,
+  "lora_bias": false,
+  "lora_dropout": 0.05,
+  "lora_ga_config": null,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "monteclora_config": null,
+  "peft_type": "LORA",
+  "peft_version": "0.20.0",
+  "qalora_group_size": 16,
+  "r": 16,
+  "rank_pattern": {},
+  "revision": "1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0",
+  "target_modules": ".*(?:language_model\\.)?layers\\.\\d+\\.(?:self_attn\\.(?:q_proj|k_proj|v_proj|o_proj)|linear_attn\\.(?:in_proj_qkv|in_proj_z|in_proj_a|in_proj_b|out_proj)|mlp\\.(?:gate_proj|up_proj|down_proj))$",
+  "target_parameters": null,
+  "task_type": "CAUSAL_LM",
+  "trainable_token_indices": null,
+  "use_bdlora": null,
+  "use_dora": false,
+  "use_qalora": false,
+  "use_rslora": false,
+  "velora_config": null
+}

training/checkpoints/checkpoint-100/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ac0fc8a29bdf8bd0de70dd275333cd6cd82e328a161ea86ac44e5e956bb85e67
+size 467047680

training/checkpoints/checkpoint-100/scheduler.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:61f4fb0f34daf13abf304a22501429019f5b65ec145513f2a2ded5685f8e4142
+size 1465

training/checkpoints/checkpoint-100/trainer_state.json

@@ -0,0 +1,734 @@
+{
+  "best_global_step": null,
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 0.18801410105757932,
+  "eval_steps": 500,
+  "global_step": 100,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 0.0018801410105757932,
+      "grad_norm": 1.1068099737167358,
+      "learning_rate": 0.0,
+      "loss": 0.18984749913215637,
+      "step": 1
+    },
+    {
+      "epoch": 0.0037602820211515863,
+      "grad_norm": 0.8424890041351318,
+      "learning_rate": 7.407407407407407e-07,
+      "loss": 0.210465207695961,
+      "step": 2
+    },
+    {
+      "epoch": 0.00564042303172738,
+      "grad_norm": 1.3556126356124878,
+      "learning_rate": 1.4814814814814815e-06,
+      "loss": 0.14603234827518463,
+      "step": 3
+    },
+    {
+      "epoch": 0.007520564042303173,
+      "grad_norm": 0.9265269637107849,
+      "learning_rate": 2.222222222222222e-06,
+      "loss": 0.17247749865055084,
+      "step": 4
+    },
+    {
+      "epoch": 0.009400705052878966,
+      "grad_norm": 0.9704127907752991,
+      "learning_rate": 2.962962962962963e-06,
+      "loss": 0.21077993512153625,
+      "step": 5
+    },
+    {
+      "epoch": 0.01128084606345476,
+      "grad_norm": 0.5968538522720337,
+      "learning_rate": 3.7037037037037037e-06,
+      "loss": 0.17516383528709412,
+      "step": 6
+    },
+    {
+      "epoch": 0.013160987074030552,
+      "grad_norm": 0.763954222202301,
+      "learning_rate": 4.444444444444444e-06,
+      "loss": 0.18361800909042358,
+      "step": 7
+    },
+    {
+      "epoch": 0.015041128084606345,
+      "grad_norm": 0.7724475264549255,
+      "learning_rate": 5.185185185185185e-06,
+      "loss": 0.23388183116912842,
+      "step": 8
+    },
+    {
+      "epoch": 0.01692126909518214,
+      "grad_norm": 0.9471437931060791,
+      "learning_rate": 5.925925925925926e-06,
+      "loss": 0.1946387141942978,
+      "step": 9
+    },
+    {
+      "epoch": 0.01880141010575793,
+      "grad_norm": 0.6618597507476807,
+      "learning_rate": 6.666666666666667e-06,
+      "loss": 0.17781661450862885,
+      "step": 10
+    },
+    {
+      "epoch": 0.020681551116333723,
+      "grad_norm": 0.8230735659599304,
+      "learning_rate": 7.4074074074074075e-06,
+      "loss": 0.23627348244190216,
+      "step": 11
+    },
+    {
+      "epoch": 0.02256169212690952,
+      "grad_norm": 0.7146603465080261,
+      "learning_rate": 8.148148148148148e-06,
+      "loss": 0.1887625753879547,
+      "step": 12
+    },
+    {
+      "epoch": 0.02444183313748531,
+      "grad_norm": 0.4817526936531067,
+      "learning_rate": 8.888888888888888e-06,
+      "loss": 0.12693078815937042,
+      "step": 13
+    },
+    {
+      "epoch": 0.026321974148061103,
+      "grad_norm": 0.6374757885932922,
+      "learning_rate": 9.62962962962963e-06,
+      "loss": 0.2233470380306244,
+      "step": 14
+    },
+    {
+      "epoch": 0.0282021151586369,
+      "grad_norm": 0.5233004093170166,
+      "learning_rate": 1.037037037037037e-05,
+      "loss": 0.1184539720416069,
+      "step": 15
+    },
+    {
+      "epoch": 0.03008225616921269,
+      "grad_norm": 0.3556949496269226,
+      "learning_rate": 1.1111111111111113e-05,
+      "loss": 0.13150134682655334,
+      "step": 16
+    },
+    {
+      "epoch": 0.031962397179788486,
+      "grad_norm": 0.3290539085865021,
+      "learning_rate": 1.1851851851851852e-05,
+      "loss": 0.15358321368694305,
+      "step": 17
+    },
+    {
+      "epoch": 0.03384253819036428,
+      "grad_norm": 0.4063560664653778,
+      "learning_rate": 1.2592592592592593e-05,
+      "loss": 0.10221662372350693,
+      "step": 18
+    },
+    {
+      "epoch": 0.03572267920094007,
+      "grad_norm": 0.3550509512424469,
+      "learning_rate": 1.3333333333333333e-05,
+      "loss": 0.1078157126903534,
+      "step": 19
+    },
+    {
+      "epoch": 0.03760282021151586,
+      "grad_norm": 0.6627258062362671,
+      "learning_rate": 1.4074074074074075e-05,
+      "loss": 0.21149909496307373,
+      "step": 20
+    },
+    {
+      "epoch": 0.039482961222091655,
+      "grad_norm": 0.46360066533088684,
+      "learning_rate": 1.4814814814814815e-05,
+      "loss": 0.1677466630935669,
+      "step": 21
+    },
+    {
+      "epoch": 0.04136310223266745,
+      "grad_norm": 0.3896161913871765,
+      "learning_rate": 1.555555555555556e-05,
+      "loss": 0.12463660538196564,
+      "step": 22
+    },
+    {
+      "epoch": 0.043243243243243246,
+      "grad_norm": 0.8209279775619507,
+      "learning_rate": 1.6296296296296297e-05,
+      "loss": 0.15096110105514526,
+      "step": 23
+    },
+    {
+      "epoch": 0.04512338425381904,
+      "grad_norm": 0.32191896438598633,
+      "learning_rate": 1.7037037037037038e-05,
+      "loss": 0.1416177600622177,
+      "step": 24
+    },
+    {
+      "epoch": 0.04700352526439483,
+      "grad_norm": 1.9557298421859741,
+      "learning_rate": 1.7777777777777777e-05,
+      "loss": 0.12210831791162491,
+      "step": 25
+    },
+    {
+      "epoch": 0.04888366627497062,
+      "grad_norm": 0.5610401630401611,
+      "learning_rate": 1.851851851851852e-05,
+      "loss": 0.13184665143489838,
+      "step": 26
+    },
+    {
+      "epoch": 0.050763807285546414,
+      "grad_norm": 0.2798151671886444,
+      "learning_rate": 1.925925925925926e-05,
+      "loss": 0.1487886756658554,
+      "step": 27
+    },
+    {
+      "epoch": 0.052643948296122206,
+      "grad_norm": 0.5388768911361694,
+      "learning_rate": 2e-05,
+      "loss": 0.15049003064632416,
+      "step": 28
+    },
+    {
+      "epoch": 0.054524089306698005,
+      "grad_norm": 0.5584218502044678,
+      "learning_rate": 1.9999806497939984e-05,
+      "loss": 0.11046376824378967,
+      "step": 29
+    },
+    {
+      "epoch": 0.0564042303172738,
+      "grad_norm": 0.4847019612789154,
+      "learning_rate": 1.999922599924854e-05,
+      "loss": 0.15236796438694,
+      "step": 30
+    },
+    {
+      "epoch": 0.05828437132784959,
+      "grad_norm": 0.30841490626335144,
+      "learning_rate": 1.9998258526391207e-05,
+      "loss": 0.10093137621879578,
+      "step": 31
+    },
+    {
+      "epoch": 0.06016451233842538,
+      "grad_norm": 0.39490729570388794,
+      "learning_rate": 1.9996904116809586e-05,
+      "loss": 0.15531431138515472,
+      "step": 32
+    },
+    {
+      "epoch": 0.06204465334900117,
+      "grad_norm": 0.5662699937820435,
+      "learning_rate": 1.999516282291988e-05,
+      "loss": 0.21666380763053894,
+      "step": 33
+    },
+    {
+      "epoch": 0.06392479435957697,
+      "grad_norm": 0.42729634046554565,
+      "learning_rate": 1.9993034712110888e-05,
+      "loss": 0.12972058355808258,
+      "step": 34
+    },
+    {
+      "epoch": 0.06580493537015276,
+      "grad_norm": 0.35263097286224365,
+      "learning_rate": 1.999051986674137e-05,
+      "loss": 0.12465275079011917,
+      "step": 35
+    },
+    {
+      "epoch": 0.06768507638072856,
+      "grad_norm": 0.535301685333252,
+      "learning_rate": 1.998761838413688e-05,
+      "loss": 0.09896158427000046,
+      "step": 36
+    },
+    {
+      "epoch": 0.06956521739130435,
+      "grad_norm": 0.3752855062484741,
+      "learning_rate": 1.998433037658599e-05,
+      "loss": 0.1280779242515564,
+      "step": 37
+    },
+    {
+      "epoch": 0.07144535840188014,
+      "grad_norm": 0.981087327003479,
+      "learning_rate": 1.9980655971335944e-05,
+      "loss": 0.09660309553146362,
+      "step": 38
+    },
+    {
+      "epoch": 0.07332549941245593,
+      "grad_norm": 0.335188090801239,
+      "learning_rate": 1.997659531058774e-05,
+      "loss": 0.11915186047554016,
+      "step": 39
+    },
+    {
+      "epoch": 0.07520564042303172,
+      "grad_norm": 0.42213037610054016,
+      "learning_rate": 1.997214855149063e-05,
+      "loss": 0.11343645304441452,
+      "step": 40
+    },
+    {
+      "epoch": 0.07708578143360752,
+      "grad_norm": 0.33926352858543396,
+      "learning_rate": 1.996731586613601e-05,
+      "loss": 0.13173750042915344,
+      "step": 41
+    },
+    {
+      "epoch": 0.07896592244418331,
+      "grad_norm": 0.27011075615882874,
+      "learning_rate": 1.9962097441550803e-05,
+      "loss": 0.11818286776542664,
+      "step": 42
+    },
+    {
+      "epoch": 0.0808460634547591,
+      "grad_norm": 0.23559340834617615,
+      "learning_rate": 1.995649347969019e-05,
+      "loss": 0.05154850333929062,
+      "step": 43
+    },
+    {
+      "epoch": 0.0827262044653349,
+      "grad_norm": 0.4766392707824707,
+      "learning_rate": 1.9950504197429796e-05,
+      "loss": 0.15302187204360962,
+      "step": 44
+    },
+    {
+      "epoch": 0.0846063454759107,
+      "grad_norm": 0.49355629086494446,
+      "learning_rate": 1.9944129826557322e-05,
+      "loss": 0.13926656544208527,
+      "step": 45
+    },
+    {
+      "epoch": 0.08648648648648649,
+      "grad_norm": 0.34948262572288513,
+      "learning_rate": 1.9937370613763544e-05,
+      "loss": 0.10971198230981827,
+      "step": 46
+    },
+    {
+      "epoch": 0.08836662749706228,
+      "grad_norm": 0.42241406440734863,
+      "learning_rate": 1.993022682063278e-05,
+      "loss": 0.12660646438598633,
+      "step": 47
+    },
+    {
+      "epoch": 0.09024676850763808,
+      "grad_norm": 0.5842728018760681,
+      "learning_rate": 1.992269872363277e-05,
+      "loss": 0.11435609310865402,
+      "step": 48
+    },
+    {
+      "epoch": 0.09212690951821387,
+      "grad_norm": 0.6219758987426758,
+      "learning_rate": 1.991478661410396e-05,
+      "loss": 0.1371983140707016,
+      "step": 49
+    },
+    {
+      "epoch": 0.09400705052878966,
+      "grad_norm": 0.6244561076164246,
+      "learning_rate": 1.9906490798248267e-05,
+      "loss": 0.14317655563354492,
+      "step": 50
+    },
+    {
+      "epoch": 0.09588719153936545,
+      "grad_norm": 0.4238940477371216,
+      "learning_rate": 1.9897811597117168e-05,
+      "loss": 0.08538961410522461,
+      "step": 51
+    },
+    {
+      "epoch": 0.09776733254994124,
+      "grad_norm": 0.4316694140434265,
+      "learning_rate": 1.988874934659932e-05,
+      "loss": 0.11634167283773422,
+      "step": 52
+    },
+    {
+      "epoch": 0.09964747356051704,
+      "grad_norm": 0.27498969435691833,
+      "learning_rate": 1.987930439740757e-05,
+      "loss": 0.05380123108625412,
+      "step": 53
+    },
+    {
+      "epoch": 0.10152761457109283,
+      "grad_norm": 0.30491945147514343,
+      "learning_rate": 1.9869477115065327e-05,
+      "loss": 0.09966330975294113,
+      "step": 54
+    },
+    {
+      "epoch": 0.10340775558166862,
+      "grad_norm": 0.2556721866130829,
+      "learning_rate": 1.985926787989247e-05,
+      "loss": 0.05556664243340492,
+      "step": 55
+    },
+    {
+      "epoch": 0.10528789659224441,
+      "grad_norm": 0.29667940735816956,
+      "learning_rate": 1.9848677086990605e-05,
+      "loss": 0.053319964557886124,
+      "step": 56
+    },
+    {
+      "epoch": 0.10716803760282022,
+      "grad_norm": 0.981617271900177,
+      "learning_rate": 1.9837705146227785e-05,
+      "loss": 0.08788257837295532,
+      "step": 57
+    },
+    {
+      "epoch": 0.10904817861339601,
+      "grad_norm": 0.46110790967941284,
+      "learning_rate": 1.982635248222264e-05,
+      "loss": 0.09188224375247955,
+      "step": 58
+    },
+    {
+      "epoch": 0.1109283196239718,
+      "grad_norm": 0.998420000076294,
+      "learning_rate": 1.9814619534327936e-05,
+      "loss": 0.07375432550907135,
+      "step": 59
+    },
+    {
+      "epoch": 0.1128084606345476,
+      "grad_norm": 0.34546563029289246,
+      "learning_rate": 1.98025067566136e-05,
+      "loss": 0.09946051985025406,
+      "step": 60
+    },
+    {
+      "epoch": 0.11468860164512339,
+      "grad_norm": 0.4086509346961975,
+      "learning_rate": 1.979001461784911e-05,
+      "loss": 0.11176551133394241,
+      "step": 61
+    },
+    {
+      "epoch": 0.11656874265569918,
+      "grad_norm": 0.42808905243873596,
+      "learning_rate": 1.977714360148539e-05,
+      "loss": 0.09157419204711914,
+      "step": 62
+    },
+    {
+      "epoch": 0.11844888366627497,
+      "grad_norm": 0.3342861831188202,
+      "learning_rate": 1.976389420563607e-05,
+      "loss": 0.10024617612361908,
+      "step": 63
+    },
+    {
+      "epoch": 0.12032902467685076,
+      "grad_norm": 0.557330846786499,
+      "learning_rate": 1.975026694305824e-05,
+      "loss": 0.1302940845489502,
+      "step": 64
+    },
+    {
+      "epoch": 0.12220916568742655,
+      "grad_norm": 0.500476062297821,
+      "learning_rate": 1.9736262341132564e-05,
+      "loss": 0.10337391495704651,
+      "step": 65
+    },
+    {
+      "epoch": 0.12408930669800235,
+      "grad_norm": 0.34659484028816223,
+      "learning_rate": 1.9721880941842914e-05,
+      "loss": 0.07025514543056488,
+      "step": 66
+    },
+    {
+      "epoch": 0.12596944770857815,
+      "grad_norm": 0.5349281430244446,
+      "learning_rate": 1.970712330175536e-05,
+      "loss": 0.13431039452552795,
+      "step": 67
+    },
+    {
+      "epoch": 0.12784958871915394,
+      "grad_norm": 0.569200873374939,
+      "learning_rate": 1.9691989991996663e-05,
+      "loss": 0.11962932348251343,
+      "step": 68
+    },
+    {
+      "epoch": 0.12972972972972974,
+      "grad_norm": 1.6833670139312744,
+      "learning_rate": 1.967648159823214e-05,
+      "loss": 0.07645167410373688,
+      "step": 69
+    },
+    {
+      "epoch": 0.13160987074030553,
+      "grad_norm": 0.3905375301837921,
+      "learning_rate": 1.9660598720643018e-05,
+      "loss": 0.04278218373656273,
+      "step": 70
+    },
+    {
+      "epoch": 0.13349001175088132,
+      "grad_norm": 0.4202016294002533,
+      "learning_rate": 1.964434197390321e-05,
+      "loss": 0.1238885372877121,
+      "step": 71
+    },
+    {
+      "epoch": 0.1353701527614571,
+      "grad_norm": 0.37672147154808044,
+      "learning_rate": 1.9627711987155506e-05,
+      "loss": 0.11250915378332138,
+      "step": 72
+    },
+    {
+      "epoch": 0.1372502937720329,
+      "grad_norm": 0.41024649143218994,
+      "learning_rate": 1.9610709403987248e-05,
+      "loss": 0.04700092971324921,
+      "step": 73
+    },
+    {
+      "epoch": 0.1391304347826087,
+      "grad_norm": 0.4298761188983917,
+      "learning_rate": 1.959333488240541e-05,
+      "loss": 0.09124750643968582,
+      "step": 74
+    },
+    {
+      "epoch": 0.1410105757931845,
+      "grad_norm": 0.43265625834465027,
+      "learning_rate": 1.9575589094811128e-05,
+      "loss": 0.06408238410949707,
+      "step": 75
+    },
+    {
+      "epoch": 0.14289071680376028,
+      "grad_norm": 0.2794874906539917,
+      "learning_rate": 1.955747272797371e-05,
+      "loss": 0.057743970304727554,
+      "step": 76
+    },
+    {
+      "epoch": 0.14477085781433607,
+      "grad_norm": 0.3569641411304474,
+      "learning_rate": 1.9538986483004e-05,
+      "loss": 0.07405605912208557,
+      "step": 77
+    },
+    {
+      "epoch": 0.14665099882491187,
+      "grad_norm": 1.7423436641693115,
+      "learning_rate": 1.95201310753273e-05,
+      "loss": 0.07177209109067917,
+      "step": 78
+    },
+    {
+      "epoch": 0.14853113983548766,
+      "grad_norm": 0.5717846155166626,
+      "learning_rate": 1.950090723465566e-05,
+      "loss": 0.11345040798187256,
+      "step": 79
+    },
+    {
+      "epoch": 0.15041128084606345,
+      "grad_norm": 0.47322016954421997,
+      "learning_rate": 1.948131570495963e-05,
+      "loss": 0.05455246940255165,
+      "step": 80
+    },
+    {
+      "epoch": 0.15229142185663924,
+      "grad_norm": 0.5463703274726868,
+      "learning_rate": 1.946135724443948e-05,
+      "loss": 0.12550315260887146,
+      "step": 81
+    },
+    {
+      "epoch": 0.15417156286721503,
+      "grad_norm": 0.4568125009536743,
+      "learning_rate": 1.944103262549586e-05,
+      "loss": 0.0811210423707962,
+      "step": 82
+    },
+    {
+      "epoch": 0.15605170387779083,
+      "grad_norm": 0.8583007454872131,
+      "learning_rate": 1.9420342634699893e-05,
+      "loss": 0.20053160190582275,
+      "step": 83
+    },
+    {
+      "epoch": 0.15793184488836662,
+      "grad_norm": 0.40888386964797974,
+      "learning_rate": 1.939928807276275e-05,
+      "loss": 0.07020855695009232,
+      "step": 84
+    },
+    {
+      "epoch": 0.1598119858989424,
+      "grad_norm": 0.5354244112968445,
+      "learning_rate": 1.9377869754504645e-05,
+      "loss": 0.07153885811567307,
+      "step": 85
+    },
+    {
+      "epoch": 0.1616921269095182,
+      "grad_norm": 0.8294443488121033,
+      "learning_rate": 1.935608850882333e-05,
+      "loss": 0.07079675048589706,
+      "step": 86
+    },
+    {
+      "epoch": 0.163572267920094,
+      "grad_norm": 0.3602754771709442,
+      "learning_rate": 1.933394517866198e-05,
+      "loss": 0.04375575855374336,
+      "step": 87
+    },
+    {
+      "epoch": 0.1654524089306698,
+      "grad_norm": 0.7854155898094177,
+      "learning_rate": 1.9311440620976597e-05,
+      "loss": 0.05288895219564438,
+      "step": 88
+    },
+    {
+      "epoch": 0.1673325499412456,
+      "grad_norm": 0.36262285709381104,
+      "learning_rate": 1.9288575706702836e-05,
+      "loss": 0.05213308706879616,
+      "step": 89
+    },
+    {
+      "epoch": 0.1692126909518214,
+      "grad_norm": 0.7056986093521118,
+      "learning_rate": 1.9265351320722296e-05,
+      "loss": 0.08565079420804977,
+      "step": 90
+    },
+    {
+      "epoch": 0.1710928319623972,
+      "grad_norm": 0.390341192483902,
+      "learning_rate": 1.924176836182829e-05,
+      "loss": 0.0380721278488636,
+      "step": 91
+    },
+    {
+      "epoch": 0.17297297297297298,
+      "grad_norm": 0.9817848205566406,
+      "learning_rate": 1.921782774269104e-05,
+      "loss": 0.14400216937065125,
+      "step": 92
+    },
+    {
+      "epoch": 0.17485311398354877,
+      "grad_norm": 0.4562535285949707,
+      "learning_rate": 1.9193530389822364e-05,
+      "loss": 0.07196680456399918,
+      "step": 93
+    },
+    {
+      "epoch": 0.17673325499412457,
+      "grad_norm": 0.5924327373504639,
+      "learning_rate": 1.916887724353984e-05,
+      "loss": 0.08313235640525818,
+      "step": 94
+    },
+    {
+      "epoch": 0.17861339600470036,
+      "grad_norm": 0.3936920464038849,
+      "learning_rate": 1.914386925793038e-05,
+      "loss": 0.08207011967897415,
+      "step": 95
+    },
+    {
+      "epoch": 0.18049353701527615,
+      "grad_norm": 0.4106353521347046,
+      "learning_rate": 1.9118507400813327e-05,
+      "loss": 0.07451261579990387,
+      "step": 96
+    },
+    {
+      "epoch": 0.18237367802585194,
+      "grad_norm": 0.8423656225204468,
+      "learning_rate": 1.9092792653703006e-05,
+      "loss": 0.1120993122458458,
+      "step": 97
+    },
+    {
+      "epoch": 0.18425381903642774,
+      "grad_norm": 0.4818660616874695,
+      "learning_rate": 1.9066726011770725e-05,
+      "loss": 0.0911557748913765,
+      "step": 98
+    },
+    {
+      "epoch": 0.18613396004700353,
+      "grad_norm": 1.91852867603302,
+      "learning_rate": 1.9040308483806266e-05,
+      "loss": 0.08450711518526077,
+      "step": 99
+    },
+    {
+      "epoch": 0.18801410105757932,
+      "grad_norm": 0.5150951147079468,
+      "learning_rate": 1.9013541092178844e-05,
+      "loss": 0.07893037796020508,
+      "step": 100
+    }
+  ],
+  "logging_steps": 1,
+  "max_steps": 532,
+  "num_input_tokens_seen": 0,
+  "num_train_epochs": 1,
+  "save_steps": 100,
+  "stateful_callbacks": {
+    "TrainerControl": {
+      "args": {
+        "should_epoch_stop": false,
+        "should_evaluate": false,
+        "should_log": false,
+        "should_save": true,
+        "should_training_stop": false
+      },
+      "attributes": {}
+    }
+  },
+  "total_flos": 1.684346206965627e+17,
+  "train_batch_size": 1,
+  "trial_name": null,
+  "trial_params": null
+}

training/checkpoints/checkpoint-200/CHECKSUMS.sha256

@@ -0,0 +1,8 @@
+df9859f741f75a96a958b358a1c08817e5fa6ff4a8fe8eed85a01add501ae7f1  README.md
+a928e2f4694e2145d7e3c6c7f892bcdc9591a45baa6fa97e380ccb01d495b875  adapter_config.json
+8a2dd645a8a3acb7b698820dd0538c355a193825a29513eed313622dcedb16b1  adapter_model.safetensors
+0068df00f7ef212f3431f642c2951d59ee4941f490ab606d14e495b23565c097  optimizer.pt
+7a15f7f8f5cb1ba94601bba78e166ad271e3baa01d4ed9bdbec9d4d9f03db013  rng_state.pth
+532c3998af11254097768cba2a02356701312926356a1ae080ad6f6e41462901  scheduler.pt
+0f0e0df58888e81b7d798a7e444b81fe3df1e548ee1200c602a94e5a6eb29651  trainer_state.json
+175111135c9d40b63054d1ef321345f80de8c4d21717241624ae93bced357d6c  training_args.bin

training/checkpoints/checkpoint-200/adapter_config.json

@@ -0,0 +1,42 @@
+{
+  "alora_invocation_tokens": null,
+  "alpha_pattern": {},
+  "arrow_config": null,
+  "auto_mapping": null,
+  "base_model_name_or_path": "Qwen/Qwen3.8-27B",
+  "bias": "none",
+  "corda_config": null,
+  "ensure_weight_tying": false,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 32,
+  "lora_bias": false,
+  "lora_dropout": 0.05,
+  "lora_ga_config": null,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "monteclora_config": null,
+  "peft_type": "LORA",
+  "peft_version": "0.20.0",
+  "qalora_group_size": 16,
+  "r": 16,
+  "rank_pattern": {},
+  "revision": "1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0",
+  "target_modules": ".*(?:language_model\\.)?layers\\.\\d+\\.(?:self_attn\\.(?:q_proj|k_proj|v_proj|o_proj)|linear_attn\\.(?:in_proj_qkv|in_proj_z|in_proj_a|in_proj_b|out_proj)|mlp\\.(?:gate_proj|up_proj|down_proj))$",
+  "target_parameters": null,
+  "task_type": "CAUSAL_LM",
+  "trainable_token_indices": null,
+  "use_bdlora": null,
+  "use_dora": false,
+  "use_qalora": false,
+  "use_rslora": false,
+  "velora_config": null
+}

training/checkpoints/checkpoint-200/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8a2dd645a8a3acb7b698820dd0538c355a193825a29513eed313622dcedb16b1
+size 467047680

training/checkpoints/checkpoint-200/scheduler.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:532c3998af11254097768cba2a02356701312926356a1ae080ad6f6e41462901
+size 1465

training/checkpoints/checkpoint-200/trainer_state.json

@@ -0,0 +1,1434 @@
+{
+  "best_global_step": null,
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 0.37602820211515864,
+  "eval_steps": 500,
+  "global_step": 200,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 0.0018801410105757932,
+      "grad_norm": 1.1068099737167358,
+      "learning_rate": 0.0,
+      "loss": 0.18984749913215637,
+      "step": 1
+    },
+    {
+      "epoch": 0.0037602820211515863,
+      "grad_norm": 0.8424890041351318,
+      "learning_rate": 7.407407407407407e-07,
+      "loss": 0.210465207695961,
+      "step": 2
+    },
+    {
+      "epoch": 0.00564042303172738,
+      "grad_norm": 1.3556126356124878,
+      "learning_rate": 1.4814814814814815e-06,
+      "loss": 0.14603234827518463,
+      "step": 3
+    },
+    {
+      "epoch": 0.007520564042303173,
+      "grad_norm": 0.9265269637107849,
+      "learning_rate": 2.222222222222222e-06,
+      "loss": 0.17247749865055084,
+      "step": 4
+    },
+    {
+      "epoch": 0.009400705052878966,
+      "grad_norm": 0.9704127907752991,
+      "learning_rate": 2.962962962962963e-06,
+      "loss": 0.21077993512153625,
+      "step": 5
+    },
+    {
+      "epoch": 0.01128084606345476,
+      "grad_norm": 0.5968538522720337,
+      "learning_rate": 3.7037037037037037e-06,
+      "loss": 0.17516383528709412,
+      "step": 6
+    },
+    {
+      "epoch": 0.013160987074030552,
+      "grad_norm": 0.763954222202301,
+      "learning_rate": 4.444444444444444e-06,
+      "loss": 0.18361800909042358,
+      "step": 7
+    },
+    {
+      "epoch": 0.015041128084606345,
+      "grad_norm": 0.7724475264549255,
+      "learning_rate": 5.185185185185185e-06,
+      "loss": 0.23388183116912842,
+      "step": 8
+    },
+    {
+      "epoch": 0.01692126909518214,
+      "grad_norm": 0.9471437931060791,
+      "learning_rate": 5.925925925925926e-06,
+      "loss": 0.1946387141942978,
+      "step": 9
+    },
+    {
+      "epoch": 0.01880141010575793,
+      "grad_norm": 0.6618597507476807,
+      "learning_rate": 6.666666666666667e-06,
+      "loss": 0.17781661450862885,
+      "step": 10
+    },
+    {
+      "epoch": 0.020681551116333723,
+      "grad_norm": 0.8230735659599304,
+      "learning_rate": 7.4074074074074075e-06,
+      "loss": 0.23627348244190216,
+      "step": 11
+    },
+    {
+      "epoch": 0.02256169212690952,
+      "grad_norm": 0.7146603465080261,
+      "learning_rate": 8.148148148148148e-06,
+      "loss": 0.1887625753879547,
+      "step": 12
+    },
+    {
+      "epoch": 0.02444183313748531,
+      "grad_norm": 0.4817526936531067,
+      "learning_rate": 8.888888888888888e-06,
+      "loss": 0.12693078815937042,
+      "step": 13
+    },
+    {
+      "epoch": 0.026321974148061103,
+      "grad_norm": 0.6374757885932922,
+      "learning_rate": 9.62962962962963e-06,
+      "loss": 0.2233470380306244,
+      "step": 14
+    },
+    {
+      "epoch": 0.0282021151586369,
+      "grad_norm": 0.5233004093170166,
+      "learning_rate": 1.037037037037037e-05,
+      "loss": 0.1184539720416069,
+      "step": 15
+    },
+    {
+      "epoch": 0.03008225616921269,
+      "grad_norm": 0.3556949496269226,
+      "learning_rate": 1.1111111111111113e-05,
+      "loss": 0.13150134682655334,
+      "step": 16
+    },
+    {
+      "epoch": 0.031962397179788486,
+      "grad_norm": 0.3290539085865021,
+      "learning_rate": 1.1851851851851852e-05,
+      "loss": 0.15358321368694305,
+      "step": 17
+    },
+    {
+      "epoch": 0.03384253819036428,
+      "grad_norm": 0.4063560664653778,
+      "learning_rate": 1.2592592592592593e-05,
+      "loss": 0.10221662372350693,
+      "step": 18
+    },
+    {
+      "epoch": 0.03572267920094007,
+      "grad_norm": 0.3550509512424469,
+      "learning_rate": 1.3333333333333333e-05,
+      "loss": 0.1078157126903534,
+      "step": 19
+    },
+    {
+      "epoch": 0.03760282021151586,
+      "grad_norm": 0.6627258062362671,
+      "learning_rate": 1.4074074074074075e-05,
+      "loss": 0.21149909496307373,
+      "step": 20
+    },
+    {
+      "epoch": 0.039482961222091655,
+      "grad_norm": 0.46360066533088684,
+      "learning_rate": 1.4814814814814815e-05,
+      "loss": 0.1677466630935669,
+      "step": 21
+    },
+    {
+      "epoch": 0.04136310223266745,
+      "grad_norm": 0.3896161913871765,
+      "learning_rate": 1.555555555555556e-05,
+      "loss": 0.12463660538196564,
+      "step": 22
+    },
+    {
+      "epoch": 0.043243243243243246,
+      "grad_norm": 0.8209279775619507,
+      "learning_rate": 1.6296296296296297e-05,
+      "loss": 0.15096110105514526,
+      "step": 23
+    },
+    {
+      "epoch": 0.04512338425381904,
+      "grad_norm": 0.32191896438598633,
+      "learning_rate": 1.7037037037037038e-05,
+      "loss": 0.1416177600622177,
+      "step": 24
+    },
+    {
+      "epoch": 0.04700352526439483,
+      "grad_norm": 1.9557298421859741,
+      "learning_rate": 1.7777777777777777e-05,
+      "loss": 0.12210831791162491,
+      "step": 25
+    },
+    {
+      "epoch": 0.04888366627497062,
+      "grad_norm": 0.5610401630401611,
+      "learning_rate": 1.851851851851852e-05,
+      "loss": 0.13184665143489838,
+      "step": 26
+    },
+    {
+      "epoch": 0.050763807285546414,
+      "grad_norm": 0.2798151671886444,
+      "learning_rate": 1.925925925925926e-05,
+      "loss": 0.1487886756658554,
+      "step": 27
+    },
+    {
+      "epoch": 0.052643948296122206,
+      "grad_norm": 0.5388768911361694,
+      "learning_rate": 2e-05,
+      "loss": 0.15049003064632416,
+      "step": 28
+    },
+    {
+      "epoch": 0.054524089306698005,
+      "grad_norm": 0.5584218502044678,
+      "learning_rate": 1.9999806497939984e-05,
+      "loss": 0.11046376824378967,
+      "step": 29
+    },
+    {
+      "epoch": 0.0564042303172738,
+      "grad_norm": 0.4847019612789154,
+      "learning_rate": 1.999922599924854e-05,
+      "loss": 0.15236796438694,
+      "step": 30
+    },
+    {
+      "epoch": 0.05828437132784959,
+      "grad_norm": 0.30841490626335144,
+      "learning_rate": 1.9998258526391207e-05,
+      "loss": 0.10093137621879578,
+      "step": 31
+    },
+    {
+      "epoch": 0.06016451233842538,
+      "grad_norm": 0.39490729570388794,
+      "learning_rate": 1.9996904116809586e-05,
+      "loss": 0.15531431138515472,
+      "step": 32
+    },
+    {
+      "epoch": 0.06204465334900117,
+      "grad_norm": 0.5662699937820435,
+      "learning_rate": 1.999516282291988e-05,
+      "loss": 0.21666380763053894,
+      "step": 33
+    },
+    {
+      "epoch": 0.06392479435957697,
+      "grad_norm": 0.42729634046554565,
+      "learning_rate": 1.9993034712110888e-05,
+      "loss": 0.12972058355808258,
+      "step": 34
+    },
+    {
+      "epoch": 0.06580493537015276,
+      "grad_norm": 0.35263097286224365,
+      "learning_rate": 1.999051986674137e-05,
+      "loss": 0.12465275079011917,
+      "step": 35
+    },
+    {
+      "epoch": 0.06768507638072856,
+      "grad_norm": 0.535301685333252,
+      "learning_rate": 1.998761838413688e-05,
+      "loss": 0.09896158427000046,
+      "step": 36
+    },
+    {
+      "epoch": 0.06956521739130435,
+      "grad_norm": 0.3752855062484741,
+      "learning_rate": 1.998433037658599e-05,
+      "loss": 0.1280779242515564,
+      "step": 37
+    },
+    {
+      "epoch": 0.07144535840188014,
+      "grad_norm": 0.981087327003479,
+      "learning_rate": 1.9980655971335944e-05,
+      "loss": 0.09660309553146362,
+      "step": 38
+    },
+    {
+      "epoch": 0.07332549941245593,
+      "grad_norm": 0.335188090801239,
+      "learning_rate": 1.997659531058774e-05,
+      "loss": 0.11915186047554016,
+      "step": 39
+    },
+    {
+      "epoch": 0.07520564042303172,
+      "grad_norm": 0.42213037610054016,
+      "learning_rate": 1.997214855149063e-05,
+      "loss": 0.11343645304441452,
+      "step": 40
+    },
+    {
+      "epoch": 0.07708578143360752,
+      "grad_norm": 0.33926352858543396,
+      "learning_rate": 1.996731586613601e-05,
+      "loss": 0.13173750042915344,
+      "step": 41
+    },
+    {
+      "epoch": 0.07896592244418331,
+      "grad_norm": 0.27011075615882874,
+      "learning_rate": 1.9962097441550803e-05,
+      "loss": 0.11818286776542664,
+      "step": 42
+    },
+    {
+      "epoch": 0.0808460634547591,
+      "grad_norm": 0.23559340834617615,
+      "learning_rate": 1.995649347969019e-05,
+      "loss": 0.05154850333929062,
+      "step": 43
+    },
+    {
+      "epoch": 0.0827262044653349,
+      "grad_norm": 0.4766392707824707,
+      "learning_rate": 1.9950504197429796e-05,
+      "loss": 0.15302187204360962,
+      "step": 44
+    },
+    {
+      "epoch": 0.0846063454759107,
+      "grad_norm": 0.49355629086494446,
+      "learning_rate": 1.9944129826557322e-05,
+      "loss": 0.13926656544208527,
+      "step": 45
+    },
+    {
+      "epoch": 0.08648648648648649,
+      "grad_norm": 0.34948262572288513,
+      "learning_rate": 1.9937370613763544e-05,
+      "loss": 0.10971198230981827,
+      "step": 46
+    },
+    {
+      "epoch": 0.08836662749706228,
+      "grad_norm": 0.42241406440734863,
+      "learning_rate": 1.993022682063278e-05,
+      "loss": 0.12660646438598633,
+      "step": 47
+    },
+    {
+      "epoch": 0.09024676850763808,
+      "grad_norm": 0.5842728018760681,
+      "learning_rate": 1.992269872363277e-05,
+      "loss": 0.11435609310865402,
+      "step": 48
+    },
+    {
+      "epoch": 0.09212690951821387,
+      "grad_norm": 0.6219758987426758,
+      "learning_rate": 1.991478661410396e-05,
+      "loss": 0.1371983140707016,
+      "step": 49
+    },
+    {
+      "epoch": 0.09400705052878966,
+      "grad_norm": 0.6244561076164246,
+      "learning_rate": 1.9906490798248267e-05,
+      "loss": 0.14317655563354492,
+      "step": 50
+    },
+    {
+      "epoch": 0.09588719153936545,
+      "grad_norm": 0.4238940477371216,
+      "learning_rate": 1.9897811597117168e-05,
+      "loss": 0.08538961410522461,
+      "step": 51
+    },
+    {
+      "epoch": 0.09776733254994124,
+      "grad_norm": 0.4316694140434265,
+      "learning_rate": 1.988874934659932e-05,
+      "loss": 0.11634167283773422,
+      "step": 52
+    },
+    {
+      "epoch": 0.09964747356051704,
+      "grad_norm": 0.27498969435691833,
+      "learning_rate": 1.987930439740757e-05,
+      "loss": 0.05380123108625412,
+      "step": 53
+    },
+    {
+      "epoch": 0.10152761457109283,
+      "grad_norm": 0.30491945147514343,
+      "learning_rate": 1.9869477115065327e-05,
+      "loss": 0.09966330975294113,
+      "step": 54
+    },
+    {
+      "epoch": 0.10340775558166862,
+      "grad_norm": 0.2556721866130829,
+      "learning_rate": 1.985926787989247e-05,
+      "loss": 0.05556664243340492,
+      "step": 55
+    },
+    {
+      "epoch": 0.10528789659224441,
+      "grad_norm": 0.29667940735816956,
+      "learning_rate": 1.9848677086990605e-05,
+      "loss": 0.053319964557886124,
+      "step": 56
+    },
+    {
+      "epoch": 0.10716803760282022,
+      "grad_norm": 0.981617271900177,
+      "learning_rate": 1.9837705146227785e-05,
+      "loss": 0.08788257837295532,
+      "step": 57
+    },
+    {
+      "epoch": 0.10904817861339601,
+      "grad_norm": 0.46110790967941284,
+      "learning_rate": 1.982635248222264e-05,
+      "loss": 0.09188224375247955,
+      "step": 58
+    },
+    {
+      "epoch": 0.1109283196239718,
+      "grad_norm": 0.998420000076294,
+      "learning_rate": 1.9814619534327936e-05,
+      "loss": 0.07375432550907135,
+      "step": 59
+    },
+    {
+      "epoch": 0.1128084606345476,
+      "grad_norm": 0.34546563029289246,
+      "learning_rate": 1.98025067566136e-05,
+      "loss": 0.09946051985025406,
+      "step": 60
+    },
+    {
+      "epoch": 0.11468860164512339,
+      "grad_norm": 0.4086509346961975,
+      "learning_rate": 1.979001461784911e-05,
+      "loss": 0.11176551133394241,
+      "step": 61
+    },
+    {
+      "epoch": 0.11656874265569918,
+      "grad_norm": 0.42808905243873596,
+      "learning_rate": 1.977714360148539e-05,
+      "loss": 0.09157419204711914,
+      "step": 62
+    },
+    {
+      "epoch": 0.11844888366627497,
+      "grad_norm": 0.3342861831188202,
+      "learning_rate": 1.976389420563607e-05,
+      "loss": 0.10024617612361908,
+      "step": 63
+    },
+    {
+      "epoch": 0.12032902467685076,
+      "grad_norm": 0.557330846786499,
+      "learning_rate": 1.975026694305824e-05,
+      "loss": 0.1302940845489502,
+      "step": 64
+    },
+    {
+      "epoch": 0.12220916568742655,
+      "grad_norm": 0.500476062297821,
+      "learning_rate": 1.9736262341132564e-05,
+      "loss": 0.10337391495704651,
+      "step": 65
+    },
+    {
+      "epoch": 0.12408930669800235,
+      "grad_norm": 0.34659484028816223,
+      "learning_rate": 1.9721880941842914e-05,
+      "loss": 0.07025514543056488,
+      "step": 66
+    },
+    {
+      "epoch": 0.12596944770857815,
+      "grad_norm": 0.5349281430244446,
+      "learning_rate": 1.970712330175536e-05,
+      "loss": 0.13431039452552795,
+      "step": 67
+    },
+    {
+      "epoch": 0.12784958871915394,
+      "grad_norm": 0.569200873374939,
+      "learning_rate": 1.9691989991996663e-05,
+      "loss": 0.11962932348251343,
+      "step": 68
+    },
+    {
+      "epoch": 0.12972972972972974,
+      "grad_norm": 1.6833670139312744,
+      "learning_rate": 1.967648159823214e-05,
+      "loss": 0.07645167410373688,
+      "step": 69
+    },
+    {
+      "epoch": 0.13160987074030553,
+      "grad_norm": 0.3905375301837921,
+      "learning_rate": 1.9660598720643018e-05,
+      "loss": 0.04278218373656273,
+      "step": 70
+    },
+    {
+      "epoch": 0.13349001175088132,
+      "grad_norm": 0.4202016294002533,
+      "learning_rate": 1.964434197390321e-05,
+      "loss": 0.1238885372877121,
+      "step": 71
+    },
+    {
+      "epoch": 0.1353701527614571,
+      "grad_norm": 0.37672147154808044,
+      "learning_rate": 1.9627711987155506e-05,
+      "loss": 0.11250915378332138,
+      "step": 72
+    },
+    {
+      "epoch": 0.1372502937720329,
+      "grad_norm": 0.41024649143218994,
+      "learning_rate": 1.9610709403987248e-05,
+      "loss": 0.04700092971324921,
+      "step": 73
+    },
+    {
+      "epoch": 0.1391304347826087,
+      "grad_norm": 0.4298761188983917,
+      "learning_rate": 1.959333488240541e-05,
+      "loss": 0.09124750643968582,
+      "step": 74
+    },
+    {
+      "epoch": 0.1410105757931845,
+      "grad_norm": 0.43265625834465027,
+      "learning_rate": 1.9575589094811128e-05,
+      "loss": 0.06408238410949707,
+      "step": 75
+    },
+    {
+      "epoch": 0.14289071680376028,
+      "grad_norm": 0.2794874906539917,
+      "learning_rate": 1.955747272797371e-05,
+      "loss": 0.057743970304727554,
+      "step": 76
+    },
+    {
+      "epoch": 0.14477085781433607,
+      "grad_norm": 0.3569641411304474,
+      "learning_rate": 1.9538986483004e-05,
+      "loss": 0.07405605912208557,
+      "step": 77
+    },
+    {
+      "epoch": 0.14665099882491187,
+      "grad_norm": 1.7423436641693115,
+      "learning_rate": 1.95201310753273e-05,
+      "loss": 0.07177209109067917,
+      "step": 78
+    },
+    {
+      "epoch": 0.14853113983548766,
+      "grad_norm": 0.5717846155166626,
+      "learning_rate": 1.950090723465566e-05,
+      "loss": 0.11345040798187256,
+      "step": 79
+    },
+    {
+      "epoch": 0.15041128084606345,
+      "grad_norm": 0.47322016954421997,
+      "learning_rate": 1.948131570495963e-05,
+      "loss": 0.05455246940255165,
+      "step": 80
+    },
+    {
+      "epoch": 0.15229142185663924,
+      "grad_norm": 0.5463703274726868,
+      "learning_rate": 1.946135724443948e-05,
+      "loss": 0.12550315260887146,
+      "step": 81
+    },
+    {
+      "epoch": 0.15417156286721503,
+      "grad_norm": 0.4568125009536743,
+      "learning_rate": 1.944103262549586e-05,
+      "loss": 0.0811210423707962,
+      "step": 82
+    },
+    {
+      "epoch": 0.15605170387779083,
+      "grad_norm": 0.8583007454872131,
+      "learning_rate": 1.9420342634699893e-05,
+      "loss": 0.20053160190582275,
+      "step": 83
+    },
+    {
+      "epoch": 0.15793184488836662,
+      "grad_norm": 0.40888386964797974,
+      "learning_rate": 1.939928807276275e-05,
+      "loss": 0.07020855695009232,
+      "step": 84
+    },
+    {
+      "epoch": 0.1598119858989424,
+      "grad_norm": 0.5354244112968445,
+      "learning_rate": 1.9377869754504645e-05,
+      "loss": 0.07153885811567307,
+      "step": 85
+    },
+    {
+      "epoch": 0.1616921269095182,
+      "grad_norm": 0.8294443488121033,
+      "learning_rate": 1.935608850882333e-05,
+      "loss": 0.07079675048589706,
+      "step": 86
+    },
+    {
+      "epoch": 0.163572267920094,
+      "grad_norm": 0.3602754771709442,
+      "learning_rate": 1.933394517866198e-05,
+      "loss": 0.04375575855374336,
+      "step": 87
+    },
+    {
+      "epoch": 0.1654524089306698,
+      "grad_norm": 0.7854155898094177,
+      "learning_rate": 1.9311440620976597e-05,
+      "loss": 0.05288895219564438,
+      "step": 88
+    },
+    {
+      "epoch": 0.1673325499412456,
+      "grad_norm": 0.36262285709381104,
+      "learning_rate": 1.9288575706702836e-05,
+      "loss": 0.05213308706879616,
+      "step": 89
+    },
+    {
+      "epoch": 0.1692126909518214,
+      "grad_norm": 0.7056986093521118,
+      "learning_rate": 1.9265351320722296e-05,
+      "loss": 0.08565079420804977,
+      "step": 90
+    },
+    {
+      "epoch": 0.1710928319623972,
+      "grad_norm": 0.390341192483902,
+      "learning_rate": 1.924176836182829e-05,
+      "loss": 0.0380721278488636,
+      "step": 91
+    },
+    {
+      "epoch": 0.17297297297297298,
+      "grad_norm": 0.9817848205566406,
+      "learning_rate": 1.921782774269104e-05,
+      "loss": 0.14400216937065125,
+      "step": 92
+    },
+    {
+      "epoch": 0.17485311398354877,
+      "grad_norm": 0.4562535285949707,
+      "learning_rate": 1.9193530389822364e-05,
+      "loss": 0.07196680456399918,
+      "step": 93
+    },
+    {
+      "epoch": 0.17673325499412457,
+      "grad_norm": 0.5924327373504639,
+      "learning_rate": 1.916887724353984e-05,
+      "loss": 0.08313235640525818,
+      "step": 94
+    },
+    {
+      "epoch": 0.17861339600470036,
+      "grad_norm": 0.3936920464038849,
+      "learning_rate": 1.914386925793038e-05,
+      "loss": 0.08207011967897415,
+      "step": 95
+    },
+    {
+      "epoch": 0.18049353701527615,
+      "grad_norm": 0.4106353521347046,
+      "learning_rate": 1.9118507400813327e-05,
+      "loss": 0.07451261579990387,
+      "step": 96
+    },
+    {
+      "epoch": 0.18237367802585194,
+      "grad_norm": 0.8423656225204468,
+      "learning_rate": 1.9092792653703006e-05,
+      "loss": 0.1120993122458458,
+      "step": 97
+    },
+    {
+      "epoch": 0.18425381903642774,
+      "grad_norm": 0.4818660616874695,
+      "learning_rate": 1.9066726011770725e-05,
+      "loss": 0.0911557748913765,
+      "step": 98
+    },
+    {
+      "epoch": 0.18613396004700353,
+      "grad_norm": 1.91852867603302,
+      "learning_rate": 1.9040308483806266e-05,
+      "loss": 0.08450711518526077,
+      "step": 99
+    },
+    {
+      "epoch": 0.18801410105757932,
+      "grad_norm": 0.5150951147079468,
+      "learning_rate": 1.9013541092178844e-05,
+      "loss": 0.07893037796020508,
+      "step": 100
+    },
+    {
+      "epoch": 0.1898942420681551,
+      "grad_norm": 0.43682360649108887,
+      "learning_rate": 1.8986424872797542e-05,
+      "loss": 0.07814378291368484,
+      "step": 101
+    },
+    {
+      "epoch": 0.1917743830787309,
+      "grad_norm": 1.8135225772857666,
+      "learning_rate": 1.8958960875071225e-05,
+      "loss": 0.12606708705425262,
+      "step": 102
+    },
+    {
+      "epoch": 0.1936545240893067,
+      "grad_norm": 0.29491138458251953,
+      "learning_rate": 1.8931150161867917e-05,
+      "loss": 0.040284134447574615,
+      "step": 103
+    },
+    {
+      "epoch": 0.1955346650998825,
+      "grad_norm": 0.3637630045413971,
+      "learning_rate": 1.890299380947368e-05,
+      "loss": 0.0349123440682888,
+      "step": 104
+    },
+    {
+      "epoch": 0.19741480611045828,
+      "grad_norm": 0.48362934589385986,
+      "learning_rate": 1.887449290755095e-05,
+      "loss": 0.07024738192558289,
+      "step": 105
+    },
+    {
+      "epoch": 0.19929494712103407,
+      "grad_norm": 0.4112212359905243,
+      "learning_rate": 1.8845648559096378e-05,
+      "loss": 0.04214769974350929,
+      "step": 106
+    },
+    {
+      "epoch": 0.20117508813160986,
+      "grad_norm": 0.7555498480796814,
+      "learning_rate": 1.8816461880398126e-05,
+      "loss": 0.12782445549964905,
+      "step": 107
+    },
+    {
+      "epoch": 0.20305522914218566,
+      "grad_norm": 0.4474226236343384,
+      "learning_rate": 1.878693400099269e-05,
+      "loss": 0.035876866430044174,
+      "step": 108
+    },
+    {
+      "epoch": 0.20493537015276145,
+      "grad_norm": 0.5831688046455383,
+      "learning_rate": 1.8757066063621167e-05,
+      "loss": 0.0876915454864502,
+      "step": 109
+    },
+    {
+      "epoch": 0.20681551116333724,
+      "grad_norm": 0.2091899812221527,
+      "learning_rate": 1.872685922418504e-05,
+      "loss": 0.034780748188495636,
+      "step": 110
+    },
+    {
+      "epoch": 0.20869565217391303,
+      "grad_norm": 0.5429494976997375,
+      "learning_rate": 1.8696314651701438e-05,
+      "loss": 0.04790528863668442,
+      "step": 111
+    },
+    {
+      "epoch": 0.21057579318448882,
+      "grad_norm": 0.7049829959869385,
+      "learning_rate": 1.8665433528257904e-05,
+      "loss": 0.07777579128742218,
+      "step": 112
+    },
+    {
+      "epoch": 0.21245593419506462,
+      "grad_norm": 0.5686396956443787,
+      "learning_rate": 1.8634217048966638e-05,
+      "loss": 0.06032140552997589,
+      "step": 113
+    },
+    {
+      "epoch": 0.21433607520564044,
+      "grad_norm": 0.6066485047340393,
+      "learning_rate": 1.8602666421918245e-05,
+      "loss": 0.06071196123957634,
+      "step": 114
+    },
+    {
+      "epoch": 0.21621621621621623,
+      "grad_norm": 0.8486462235450745,
+      "learning_rate": 1.8570782868134997e-05,
+      "loss": 0.17299583554267883,
+      "step": 115
+    },
+    {
+      "epoch": 0.21809635722679202,
+      "grad_norm": 0.7102699279785156,
+      "learning_rate": 1.853856762152356e-05,
+      "loss": 0.09849053621292114,
+      "step": 116
+    },
+    {
+      "epoch": 0.2199764982373678,
+      "grad_norm": 0.6520931124687195,
+      "learning_rate": 1.8506021928827246e-05,
+      "loss": 0.06850472837686539,
+      "step": 117
+    },
+    {
+      "epoch": 0.2218566392479436,
+      "grad_norm": 0.5420147776603699,
+      "learning_rate": 1.8473147049577777e-05,
+      "loss": 0.04323617368936539,
+      "step": 118
+    },
+    {
+      "epoch": 0.2237367802585194,
+      "grad_norm": 0.6764035224914551,
+      "learning_rate": 1.843994425604652e-05,
+      "loss": 0.04347297176718712,
+      "step": 119
+    },
+    {
+      "epoch": 0.2256169212690952,
+      "grad_norm": 0.5113638043403625,
+      "learning_rate": 1.840641483319527e-05,
+      "loss": 0.05386022850871086,
+      "step": 120
+    },
+    {
+      "epoch": 0.22749706227967098,
+      "grad_norm": 0.4199598729610443,
+      "learning_rate": 1.8372560078626502e-05,
+      "loss": 0.04458491504192352,
+      "step": 121
+    },
+    {
+      "epoch": 0.22937720329024677,
+      "grad_norm": 0.6987719535827637,
+      "learning_rate": 1.8338381302533164e-05,
+      "loss": 0.0665862187743187,
+      "step": 122
+    },
+    {
+      "epoch": 0.23125734430082256,
+      "grad_norm": 0.6832467913627625,
+      "learning_rate": 1.8303879827647977e-05,
+      "loss": 0.061483677476644516,
+      "step": 123
+    },
+    {
+      "epoch": 0.23313748531139836,
+      "grad_norm": 0.20943143963813782,
+      "learning_rate": 1.8269056989192232e-05,
+      "loss": 0.007666974328458309,
+      "step": 124
+    },
+    {
+      "epoch": 0.23501762632197415,
+      "grad_norm": 0.9619431495666504,
+      "learning_rate": 1.823391413482412e-05,
+      "loss": 0.06872864812612534,
+      "step": 125
+    },
+    {
+      "epoch": 0.23689776733254994,
+      "grad_norm": 0.6353145837783813,
+      "learning_rate": 1.8198452624586592e-05,
+      "loss": 0.03983674198389053,
+      "step": 126
+    },
+    {
+      "epoch": 0.23877790834312573,
+      "grad_norm": 0.6979302167892456,
+      "learning_rate": 1.81626738308547e-05,
+      "loss": 0.0667322501540184,
+      "step": 127
+    },
+    {
+      "epoch": 0.24065804935370153,
+      "grad_norm": 0.5514077544212341,
+      "learning_rate": 1.8126579138282502e-05,
+      "loss": 0.046818796545267105,
+      "step": 128
+    },
+    {
+      "epoch": 0.24253819036427732,
+      "grad_norm": 0.4492374062538147,
+      "learning_rate": 1.8090169943749477e-05,
+      "loss": 0.04098084196448326,
+      "step": 129
+    },
+    {
+      "epoch": 0.2444183313748531,
+      "grad_norm": 0.6961103677749634,
+      "learning_rate": 1.8053447656306448e-05,
+      "loss": 0.04883302003145218,
+      "step": 130
+    },
+    {
+      "epoch": 0.2462984723854289,
+      "grad_norm": 0.8248452544212341,
+      "learning_rate": 1.8016413697121066e-05,
+      "loss": 0.06784651428461075,
+      "step": 131
+    },
+    {
+      "epoch": 0.2481786133960047,
+      "grad_norm": 1.0589181184768677,
+      "learning_rate": 1.797906949942282e-05,
+      "loss": 0.052728284150362015,
+      "step": 132
+    },
+    {
+      "epoch": 0.2500587544065805,
+      "grad_norm": 0.3563608229160309,
+      "learning_rate": 1.7941416508447537e-05,
+      "loss": 0.021236270666122437,
+      "step": 133
+    },
+    {
+      "epoch": 0.2519388954171563,
+      "grad_norm": 0.6616787314414978,
+      "learning_rate": 1.7903456181381486e-05,
+      "loss": 0.0754692554473877,
+      "step": 134
+    },
+    {
+      "epoch": 0.25381903642773207,
+      "grad_norm": 0.4925242066383362,
+      "learning_rate": 1.7865189987304964e-05,
+      "loss": 0.03571999445557594,
+      "step": 135
+    },
+    {
+      "epoch": 0.2556991774383079,
+      "grad_norm": 0.7398449778556824,
+      "learning_rate": 1.7826619407135444e-05,
+      "loss": 0.040718164294958115,
+      "step": 136
+    },
+    {
+      "epoch": 0.25757931844888365,
+      "grad_norm": 0.6538195610046387,
+      "learning_rate": 1.7787745933570277e-05,
+      "loss": 0.073344886302948,
+      "step": 137
+    },
+    {
+      "epoch": 0.2594594594594595,
+      "grad_norm": 0.6186607480049133,
+      "learning_rate": 1.77485710710289e-05,
+      "loss": 0.055532537400722504,
+      "step": 138
+    },
+    {
+      "epoch": 0.26133960047003524,
+      "grad_norm": 0.5715737342834473,
+      "learning_rate": 1.7709096335594632e-05,
+      "loss": 0.035070016980171204,
+      "step": 139
+    },
+    {
+      "epoch": 0.26321974148061106,
+      "grad_norm": 0.6023353934288025,
+      "learning_rate": 1.7669323254956006e-05,
+      "loss": 0.042510081082582474,
+      "step": 140
+    },
+    {
+      "epoch": 0.2650998824911868,
+      "grad_norm": 0.5790014266967773,
+      "learning_rate": 1.7629253368347624e-05,
+      "loss": 0.04061253368854523,
+      "step": 141
+    },
+    {
+      "epoch": 0.26698002350176264,
+      "grad_norm": 0.41726481914520264,
+      "learning_rate": 1.7588888226490604e-05,
+      "loss": 0.016257653012871742,
+      "step": 142
+    },
+    {
+      "epoch": 0.2688601645123384,
+      "grad_norm": 13.69245719909668,
+      "learning_rate": 1.7548229391532572e-05,
+      "loss": 0.0682084858417511,
+      "step": 143
+    },
+    {
+      "epoch": 0.2707403055229142,
+      "grad_norm": 0.3829832971096039,
+      "learning_rate": 1.7507278436987188e-05,
+      "loss": 0.029746640473604202,
+      "step": 144
+    },
+    {
+      "epoch": 0.27262044653349,
+      "grad_norm": 0.5354511737823486,
+      "learning_rate": 1.746603694767327e-05,
+      "loss": 0.008099116384983063,
+      "step": 145
+    },
+    {
+      "epoch": 0.2745005875440658,
+      "grad_norm": 0.47373664379119873,
+      "learning_rate": 1.7424506519653438e-05,
+      "loss": 0.021898245438933372,
+      "step": 146
+    },
+    {
+      "epoch": 0.2763807285546416,
+      "grad_norm": 0.5978367328643799,
+      "learning_rate": 1.7382688760172377e-05,
+      "loss": 0.0594993457198143,
+      "step": 147
+    },
+    {
+      "epoch": 0.2782608695652174,
+      "grad_norm": 0.32713404297828674,
+      "learning_rate": 1.7340585287594605e-05,
+      "loss": 0.018217433243989944,
+      "step": 148
+    },
+    {
+      "epoch": 0.28014101057579316,
+      "grad_norm": 1.1237229108810425,
+      "learning_rate": 1.729819773134185e-05,
+      "loss": 0.07925233244895935,
+      "step": 149
+    },
+    {
+      "epoch": 0.282021151586369,
+      "grad_norm": 1.1572819948196411,
+      "learning_rate": 1.725552773183001e-05,
+      "loss": 0.0854354053735733,
+      "step": 150
+    },
+    {
+      "epoch": 0.28390129259694474,
+      "grad_norm": 0.36294350028038025,
+      "learning_rate": 1.7212576940405647e-05,
+      "loss": 0.013134794309735298,
+      "step": 151
+    },
+    {
+      "epoch": 0.28578143360752056,
+      "grad_norm": 0.8305812478065491,
+      "learning_rate": 1.716934701928208e-05,
+      "loss": 0.09783323854207993,
+      "step": 152
+    },
+    {
+      "epoch": 0.2876615746180964,
+      "grad_norm": 0.6770759224891663,
+      "learning_rate": 1.7125839641475074e-05,
+      "loss": 0.06042662262916565,
+      "step": 153
+    },
+    {
+      "epoch": 0.28954171562867215,
+      "grad_norm": 0.2491256296634674,
+      "learning_rate": 1.7082056490738065e-05,
+      "loss": 0.01282151136547327,
+      "step": 154
+    },
+    {
+      "epoch": 0.29142185663924797,
+      "grad_norm": 0.3872174024581909,
+      "learning_rate": 1.7037999261497037e-05,
+      "loss": 0.032801028341054916,
+      "step": 155
+    },
+    {
+      "epoch": 0.29330199764982373,
+      "grad_norm": 0.8819592595100403,
+      "learning_rate": 1.6993669658784908e-05,
+      "loss": 0.040237247943878174,
+      "step": 156
+    },
+    {
+      "epoch": 0.29518213866039955,
+      "grad_norm": 0.5211074948310852,
+      "learning_rate": 1.6949069398175564e-05,
+      "loss": 0.046092916280031204,
+      "step": 157
+    },
+    {
+      "epoch": 0.2970622796709753,
+      "grad_norm": 1.2353624105453491,
+      "learning_rate": 1.690420020571747e-05,
+      "loss": 0.06157322973012924,
+      "step": 158
+    },
+    {
+      "epoch": 0.29894242068155114,
+      "grad_norm": 0.6376184821128845,
+      "learning_rate": 1.685906381786686e-05,
+      "loss": 0.07221483439207077,
+      "step": 159
+    },
+    {
+      "epoch": 0.3008225616921269,
+      "grad_norm": 0.4670684337615967,
+      "learning_rate": 1.681366198142054e-05,
+      "loss": 0.05798226222395897,
+      "step": 160
+    },
+    {
+      "epoch": 0.3027027027027027,
+      "grad_norm": 0.7154240012168884,
+      "learning_rate": 1.6767996453448282e-05,
+      "loss": 0.10533998161554337,
+      "step": 161
+    },
+    {
+      "epoch": 0.3045828437132785,
+      "grad_norm": 0.5200608372688293,
+      "learning_rate": 1.6722069001224842e-05,
+      "loss": 0.06703512370586395,
+      "step": 162
+    },
+    {
+      "epoch": 0.3064629847238543,
+      "grad_norm": 0.4213132858276367,
+      "learning_rate": 1.667588140216154e-05,
+      "loss": 0.04691825434565544,
+      "step": 163
+    },
+    {
+      "epoch": 0.30834312573443007,
+      "grad_norm": 0.384139746427536,
+      "learning_rate": 1.662943544373748e-05,
+      "loss": 0.024142542853951454,
+      "step": 164
+    },
+    {
+      "epoch": 0.3102232667450059,
+      "grad_norm": 0.3581804633140564,
+      "learning_rate": 1.65827329234304e-05,
+      "loss": 0.03277008235454559,
+      "step": 165
+    },
+    {
+      "epoch": 0.31210340775558165,
+      "grad_norm": 0.9509842395782471,
+      "learning_rate": 1.6535775648647074e-05,
+      "loss": 0.07664524018764496,
+      "step": 166
+    },
+    {
+      "epoch": 0.31398354876615747,
+      "grad_norm": 0.7950679063796997,
+      "learning_rate": 1.648856543665338e-05,
+      "loss": 0.04917037487030029,
+      "step": 167
+    },
+    {
+      "epoch": 0.31586368977673324,
+      "grad_norm": 0.6939692497253418,
+      "learning_rate": 1.644110411450398e-05,
+      "loss": 0.033213865011930466,
+      "step": 168
+    },
+    {
+      "epoch": 0.31774383078730906,
+      "grad_norm": 0.8434940576553345,
+      "learning_rate": 1.6393393518971587e-05,
+      "loss": 0.041798003017902374,
+      "step": 169
+    },
+    {
+      "epoch": 0.3196239717978848,
+      "grad_norm": 0.5167136192321777,
+      "learning_rate": 1.634543549647591e-05,
+      "loss": 0.03474203497171402,
+      "step": 170
+    },
+    {
+      "epoch": 0.32150411280846064,
+      "grad_norm": 0.5876272320747375,
+      "learning_rate": 1.6297231903012178e-05,
+      "loss": 0.03290121629834175,
+      "step": 171
+    },
+    {
+      "epoch": 0.3233842538190364,
+      "grad_norm": 0.7666569352149963,
+      "learning_rate": 1.6248784604079316e-05,
+      "loss": 0.05149153620004654,
+      "step": 172
+    },
+    {
+      "epoch": 0.3252643948296122,
+      "grad_norm": 0.5900854468345642,
+      "learning_rate": 1.6200095474607753e-05,
+      "loss": 0.04448067024350166,
+      "step": 173
+    },
+    {
+      "epoch": 0.327144535840188,
+      "grad_norm": 0.5376060009002686,
+      "learning_rate": 1.6151166398886862e-05,
+      "loss": 0.015022793784737587,
+      "step": 174
+    },
+    {
+      "epoch": 0.3290246768507638,
+      "grad_norm": 0.5801107287406921,
+      "learning_rate": 1.6101999270492033e-05,
+      "loss": 0.03291890025138855,
+      "step": 175
+    },
+    {
+      "epoch": 0.3309048178613396,
+      "grad_norm": 0.6087291836738586,
+      "learning_rate": 1.6052595992211388e-05,
+      "loss": 0.031155968084931374,
+      "step": 176
+    },
+    {
+      "epoch": 0.3327849588719154,
+      "grad_norm": 0.451362282037735,
+      "learning_rate": 1.6002958475972158e-05,
+      "loss": 0.009819609113037586,
+      "step": 177
+    },
+    {
+      "epoch": 0.3346650998824912,
+      "grad_norm": 0.7295314073562622,
+      "learning_rate": 1.595308864276666e-05,
+      "loss": 0.051709938794374466,
+      "step": 178
+    },
+    {
+      "epoch": 0.336545240893067,
+      "grad_norm": 0.3998778164386749,
+      "learning_rate": 1.5902988422578e-05,
+      "loss": 0.02723798155784607,
+      "step": 179
+    },
+    {
+      "epoch": 0.3384253819036428,
+      "grad_norm": 0.8522759079933167,
+      "learning_rate": 1.585265975430533e-05,
+      "loss": 0.04121178388595581,
+      "step": 180
+    },
+    {
+      "epoch": 0.34030552291421856,
+      "grad_norm": 0.5936810374259949,
+      "learning_rate": 1.5802104585688853e-05,
+      "loss": 0.06424441933631897,
+      "step": 181
+    },
+    {
+      "epoch": 0.3421856639247944,
+      "grad_norm": 0.6823294162750244,
+      "learning_rate": 1.5751324873234422e-05,
+      "loss": 0.03264414146542549,
+      "step": 182
+    },
+    {
+      "epoch": 0.34406580493537015,
+      "grad_norm": 0.3798384368419647,
+      "learning_rate": 1.570032258213783e-05,
+      "loss": 0.007725706323981285,
+      "step": 183
+    },
+    {
+      "epoch": 0.34594594594594597,
+      "grad_norm": 0.8923366665840149,
+      "learning_rate": 1.5649099686208756e-05,
+      "loss": 0.035471927374601364,
+      "step": 184
+    },
+    {
+      "epoch": 0.34782608695652173,
+      "grad_norm": 0.7076458930969238,
+      "learning_rate": 1.5597658167794374e-05,
+      "loss": 0.028757719323039055,
+      "step": 185
+    },
+    {
+      "epoch": 0.34970622796709755,
+      "grad_norm": 0.36611223220825195,
+      "learning_rate": 1.5546000017702646e-05,
+      "loss": 0.0077384174801409245,
+      "step": 186
+    },
+    {
+      "epoch": 0.3515863689776733,
+      "grad_norm": 0.5799921751022339,
+      "learning_rate": 1.549412723512526e-05,
+      "loss": 0.016421930864453316,
+      "step": 187
+    },
+    {
+      "epoch": 0.35346650998824913,
+      "grad_norm": 0.7346245646476746,
+      "learning_rate": 1.5442041827560274e-05,
+      "loss": 0.01859155297279358,
+      "step": 188
+    },
+    {
+      "epoch": 0.3553466509988249,
+      "grad_norm": 0.36168423295021057,
+      "learning_rate": 1.538974581073442e-05,
+      "loss": 0.012681198306381702,
+      "step": 189
+    },
+    {
+      "epoch": 0.3572267920094007,
+      "grad_norm": 0.40003958344459534,
+      "learning_rate": 1.5337241208525097e-05,
+      "loss": 0.019209278747439384,
+      "step": 190
+    },
+    {
+      "epoch": 0.3591069330199765,
+      "grad_norm": 0.6620081663131714,
+      "learning_rate": 1.5284530052882043e-05,
+      "loss": 0.03176989406347275,
+      "step": 191
+    },
+    {
+      "epoch": 0.3609870740305523,
+      "grad_norm": 0.3297053277492523,
+      "learning_rate": 1.5231614383748699e-05,
+      "loss": 0.012578723952174187,
+      "step": 192
+    },
+    {
+      "epoch": 0.36286721504112807,
+      "grad_norm": 0.5588814616203308,
+      "learning_rate": 1.5178496248983254e-05,
+      "loss": 0.023057013750076294,
+      "step": 193
+    },
+    {
+      "epoch": 0.3647473560517039,
+      "grad_norm": 0.5055070519447327,
+      "learning_rate": 1.512517770427942e-05,
+      "loss": 0.03681947663426399,
+      "step": 194
+    },
+    {
+      "epoch": 0.36662749706227965,
+      "grad_norm": 1.1773488521575928,
+      "learning_rate": 1.5071660813086837e-05,
+      "loss": 0.06652048230171204,
+      "step": 195
+    },
+    {
+      "epoch": 0.36850763807285547,
+      "grad_norm": 0.7722097635269165,
+      "learning_rate": 1.501794764653124e-05,
+      "loss": 0.056977998465299606,
+      "step": 196
+    },
+    {
+      "epoch": 0.37038777908343123,
+      "grad_norm": 0.6458454728126526,
+      "learning_rate": 1.4964040283334315e-05,
+      "loss": 0.02878635935485363,
+      "step": 197
+    },
+    {
+      "epoch": 0.37226792009400705,
+      "grad_norm": 1.2206151485443115,
+      "learning_rate": 1.4909940809733223e-05,
+      "loss": 0.022895192727446556,
+      "step": 198
+    },
+    {
+      "epoch": 0.3741480611045828,
+      "grad_norm": 1.1328730583190918,
+      "learning_rate": 1.4855651319399882e-05,
+      "loss": 0.023003041744232178,
+      "step": 199
+    },
+    {
+      "epoch": 0.37602820211515864,
+      "grad_norm": 0.7142350673675537,
+      "learning_rate": 1.480117391335993e-05,
+      "loss": 0.016481809318065643,
+      "step": 200
+    }
+  ],
+  "logging_steps": 1,
+  "max_steps": 532,
+  "num_input_tokens_seen": 0,
+  "num_train_epochs": 1,
+  "save_steps": 100,
+  "stateful_callbacks": {
+    "TrainerControl": {
+      "args": {
+        "should_epoch_stop": false,
+        "should_evaluate": false,
+        "should_log": false,
+        "should_save": true,
+        "should_training_stop": false
+      },
+      "attributes": {}
+    }
+  },
+  "total_flos": 3.371562057162424e+17,
+  "train_batch_size": 1,
+  "trial_name": null,
+  "trial_params": null
+}

training/checkpoints/checkpoint-300/CHECKSUMS.sha256

@@ -0,0 +1,8 @@
+df9859f741f75a96a958b358a1c08817e5fa6ff4a8fe8eed85a01add501ae7f1  README.md
+a928e2f4694e2145d7e3c6c7f892bcdc9591a45baa6fa97e380ccb01d495b875  adapter_config.json
+2a65244dee566d1ab4ab27232900310e26d237bf56fabed898b94dde78082d55  adapter_model.safetensors
+556bd2bb21f45ef11f064c93ab22296ee4d7b6f5131546d71457022cabe88899  optimizer.pt
+a7f64d5dc8fa2bdc5ae12da71cdb3395b17e4f494edf8753a032be8b167dd5d1  rng_state.pth
+841de5b0dcb7b1f112e703b0f661382afd353bc1755625a5460203340f9c49c8  scheduler.pt
+0e1ab6c3974bd751288735034a6a0e376652f6dfec572cff12d2383ca595770f  trainer_state.json
+175111135c9d40b63054d1ef321345f80de8c4d21717241624ae93bced357d6c  training_args.bin