Reaction Lens inference and evaluation bundle
Fixed structured seed23 epoch105 checkpoint for Reaction Lens Systems. The learned head ranks 15,692 Reactome reactions using frozen SPECTER2 features. This repository contains the inference bundle and cached development evaluation pack. Encoder weights are obtained separately from the pinned upstream revisions in manifest.json. File identities are recorded in the manifests.
Development cached-feature metrics: MAP 0.47698647378227255, Hit@1 0.3925115692048801, micro Recall@100 0.9015632633121642. These are development results, not independent test performance or biological validation. The evaluation pack contains 2,377 publications and 4,094 known links. Scores are uncalibrated proposals for expert review. English scientific abstracts are the intended input; general conversational queries have not been validated.
The head is MIT-licensed. Upstream model and data terms are separate; see THIRD_PARTY_NOTICES.md. This package does not include upstream encoder weights.
Files and verification
head.safetensors, catalog.jsonl, catalog-vectors.npy and manifest.json
form the inference bundle. evaluation/ contains only the development feature
arrays, offsets, labels and manifest. SHA256SUMS lists every other file in this
package. Verify it with shasum -a 256 -c SHA256SUMS from this directory.
In the accompanying Reaction Lens Systems source checkout, install with
make setup. Copy these files into artifacts/structured-seed23/, then run
make check-model for cached-feature evaluation. make encoder obtains the
pinned upstream encoder, and make serve starts the local service. Downloads
of this bundle require the source checkout's artifact manifest to identify its
actual repository and immutable revision; neither is invented in this package.
The head was fitted before this systems study; this package supports inference, not training reproduction. Its fixed logit threshold was fitted on development labels. Do not interpret threshold selection or sigmoid scores as calibrated biological probabilities. No reserved test data are included. Citation labels are incomplete and the development set was used repeatedly during development.