Dataset Viewer
Auto-converted to Parquet Duplicate
rollout_id
stringlengths
8
35
mode
stringclasses
18 values
family
stringclasses
3 values
task
stringlengths
248
494
status
stringclasses
1 value
n_turns
int32
3
20
n_tool_calls
int32
2
19
recall_episode_ids
listlengths
0
2
recall_judged_keep_ids
listlengths
0
2
final
stringlengths
50
26.6k
transcript
stringlengths
3.82k
264k
events
listlengths
6
40
r2_0906_s1
R2
R
Across our matters for client 1032 (matter ids 1032-*), find documents containing a section or provision addressing 'Start-Up Fees'. For each, name the matter and the document, state the position taken or obligation created, and compare how they differ. Cite the file path supporting every claim as (file: matters/...).
done
5
3
[]
[]
block. </think> <final>Based on a comprehensive search of all documents across matters for client 1032 (matter IDs 1032-00001 through 1032-00005), **no documents were found** containing a section or provision addressing "Start-Up Fees." I conducted two searches: 1. Exact phrase search for "Start-Up Fees" (case-insens...
=== TASK === Across our matters for client 1032 (matter ids 1032-*), find documents containing a section or provision addressing 'Start-Up Fees'. For each, name the matter and the document, state the position taken or obligation created, and compare how they differ. Cite the file path supporting every claim as (file: m...
[ { "step": 0, "role": "user", "content": "Across our matters for client 1032 (matter ids 1032-*), find documents containing a section or provision addressing 'Start-Up Fees'. For each, name the matter and the document, state the position taken or obligation created, and compare how they differ. Cite the ...
r2_1561_s0
R2
R
Across our matters for client 1002 (matter ids 1002-*), find documents containing a section or provision addressing 'Sheet: Earnout Summary'. For each, name the matter and the document, state the position taken or obligation created, and compare how they differ. Cite the file path supporting every claim as (file: matte...
done
4
3
[]
[]
No documents were found in matters/1002-* that contain a section or provision addressing 'Sheet: Earnout Summary'. Neither the exact phrase nor variations containing "Earnout" appeared in any markdown files across the 1002 matter IDs. Therefore, there are no documents to compare across matters. (file: matters/1002-*/**...
=== TASK === Across our matters for client 1002 (matter ids 1002-*), find documents containing a section or provision addressing 'Sheet: Earnout Summary'. For each, name the matter and the document, state the position taken or obligation created, and compare how they differ. Cite the file path supporting every claim as...
[ { "step": 0, "role": "user", "content": "Across our matters for client 1002 (matter ids 1002-*), find documents containing a section or provision addressing 'Sheet: Earnout Summary'. For each, name the matter and the document, state the position taken or obligation created, and compare how they differ. ...
r3_FinancialOfficerCrestlin_683_s0
R3
R
Build a dossier on Financial Officer, Crestline Packaging Solutions LLC across the firm's DMS: every matter where it appears, its role in each (client, counterparty, lender, advisor, regulator, expert, or other), and what specific actions the documents attribute to it. Check multiple matters. Cite the file path support...
done
3
2
[]
[]
**Dossier on Financial Officer, Crestline Packaging Solutions LLC** **Search Results:** No documents containing "Crestline Packaging Solutions LLC" were found in the firm's DMS. I searched all markdown files across all matters (matters/1003-*/**/*.md) using a case-insensitive grep for "Crestline" and returned no matc...
=== TASK === Build a dossier on Financial Officer, Crestline Packaging Solutions LLC across the firm's DMS: every matter where it appears, its role in each (client, counterparty, lender, advisor, regulator, expert, or other), and what specific actions the documents attribute to it. Check multiple matters. Cite the file...
[ { "step": 0, "role": "user", "content": "Build a dossier on Financial Officer, Crestline Packaging Solutions LLC across the firm's DMS: every matter where it appears, its role in each (client, counterparty, lender, advisor, regulator, expert, or other), and what specific actions the documents attribute ...
r3_UnitrinIncvAmericanGener_1617_s0
R3
R
Build a dossier on Unitrin, Inc. v. American General Corp across the firm's DMS: every matter where it appears, its role in each (client, counterparty, lender, advisor, regulator, expert, or other), and what specific actions the documents attribute to it. Check multiple matters. Cite the file path supporting every clai...
done
4
3
[]
[]
Based on my search of the Calderwood & Harkness LLP DMS, **Unitrin, Inc. v. American General Corp does not appear in any document** across the firm's document management system. I searched across all matters (1003-00001 through 1003-00005) using both "Unitrin" and "unitrin" (case-insensitive) regex patterns against al...
=== TASK === Build a dossier on Unitrin, Inc. v. American General Corp across the firm's DMS: every matter where it appears, its role in each (client, counterparty, lender, advisor, regulator, expert, or other), and what specific actions the documents attribute to it. Check multiple matters. Cite the file path supporti...
[ { "step": 0, "role": "user", "content": "Build a dossier on Unitrin, Inc. v. American General Corp across the firm's DMS: every matter where it appears, its role in each (client, counterparty, lender, advisor, regulator, expert, or other), and what specific actions the documents attribute to it. Check m...
r2_0367_s1
R2
R
Across our matters with a 'Insurance' workstream, find documents containing a section or provision addressing 'Conversion of Shares'. For each, name the matter and the document, state the position taken or obligation created, and compare how they differ. Cite the file path supporting every claim as (file: matters/...).
done
7
6
[]
[]
Based on my search across all documents in the Insurance workstream (matters/*/Insurance/**/*.md), no documents were found containing a section or provision addressing "Conversion of Shares." I searched using multiple variations including: - "Conversion of Shares" (exact phrase) - "convertible shares" - "share convers...
=== TASK === Across our matters with a 'Insurance' workstream, find documents containing a section or provision addressing 'Conversion of Shares'. For each, name the matter and the document, state the position taken or obligation created, and compare how they differ. Cite the file path supporting every claim as (file: ...
[ { "step": 0, "role": "user", "content": "Across our matters with a 'Insurance' workstream, find documents containing a section or provision addressing 'Conversion of Shares'. For each, name the matter and the document, state the position taken or obligation created, and compare how they differ. Cite the...
d6_1026-00004_rs4
D6
D
"Reconstruct the correspondence record of matter 1026-00004: read the emails and letters under matte(...TRUNCATED)
done
3
2
[]
[]
"Based on my search of the correspondence directory for matter 1026-00004, I found only one correspo(...TRUNCATED)
"=== TASK ===\nReconstruct the correspondence record of matter 1026-00004: read the emails and lette(...TRUNCATED)
[{"step":0,"role":"user","content":"Reconstruct the correspondence record of matter 1026-00004: read(...TRUNCATED)
r3_BrightmoorGroup_1172_s1
R3
R
"Build a dossier on Brightmoor Group across the firm's DMS: every matter where it appears, its role (...TRUNCATED)
done
4
3
[ "p6_r3_BrightmoorGroup_1172_s1_s0" ]
[]
"Based on a comprehensive search of the firm's DMS (matters/1003-*/**/*.md), **Brightmoor Group does(...TRUNCATED)
"=== TASK ===\nBuild a dossier on Brightmoor Group across the firm's DMS: every matter where it appe(...TRUNCATED)
[{"step":0,"role":"user","content":"Build a dossier on Brightmoor Group across the firm's DMS: every(...TRUNCATED)
r2_1693_s1
R2
R
"Across our matters for client 1034 (matter ids 1034-*), find documents containing a section or prov(...TRUNCATED)
done
4
3
[]
[]
"I searched across all matters for client 1034 (matter ids 1034-*) in the DMS for documents containi(...TRUNCATED)
"=== TASK ===\nAcross our matters for client 1034 (matter ids 1034-*), find documents containing a s(...TRUNCATED)
[{"step":0,"role":"user","content":"Across our matters for client 1034 (matter ids 1034-*), find doc(...TRUNCATED)
r3_Consultantshallpromptlyr_1976_s0
R3
R
"Build a dossier on Consultant shall promptly return to the Company across the firm's DMS: every mat(...TRUNCATED)
done
4
3
[]
[]
"Based on a comprehensive search of the entire DMS across all matters (matters/1003-00001 through ma(...TRUNCATED)
"=== TASK ===\nBuild a dossier on Consultant shall promptly return to the Company across the firm's (...TRUNCATED)
[{"step":0,"role":"user","content":"Build a dossier on Consultant shall promptly return to the Compa(...TRUNCATED)
r3_VireoGroup_1074_s0
R3
R
"Build a dossier on Vireo Group across the firm's DMS: every matter where it appears, its role in ea(...TRUNCATED)
done
4
3
[ "p6_r3_VireoGroup_1074_s0_s0", "p6_r3_VireoGroup_1074_s0_s1" ]
[]
"Based on a comprehensive search of the Calderwood & Harkness LLP (C&H) document management system, (...TRUNCATED)
"=== TASK ===\nBuild a dossier on Vireo Group across the firm's DMS: every matter where it appears, (...TRUNCATED)
[{"step":0,"role":"user","content":"Build a dossier on Vireo Group across the firm's DMS: every matt(...TRUNCATED)
End of preview. Expand in Data Studio

Seed sessions: the original trajectories behind the recall episodes (WM-RL v4)

49,017 unedited agent sessions of Qwen3.5-9B working fixed tasks over a synthetic law firm's document-management system with glob, grep and read, thinking on, temperature 0.7. This is the pool that violetxi/wmrl-v4-recall-trajectories was cut from: every recall episode keeps 1–3 of a session's tool calls, their results and its final answer byte for byte and regenerates only the thinking. The dropped calls exist only here. It is also the pool the recursive note loop's seed bank (round 0 of violetxi/wm-rl-notes-v4) was synthesized over, and the raw-trajectory slice of the earlier study's training mixtures.

  • Task families: D document reading / distillation 16,191 · R relational cross-matter search 27,890 · G grouped / ranking 4,936.
  • Ended with a <final> answer: 49,017; without one: 0.
  • 15,237 sessions have at least one recall episode derived from them (19,481 episodes in all); recall_episode_ids lists them and recall_judged_keep_ids the subset that passed the honesty judge.
  • 107 raw files the pool manifest marked unusable (no answer or no tool call) are not included.

Provenance

The underlying world is Harvey AI's synthetic legal benchmark (MIT), a 9,288-document / 266-matter law-firm corpus, revision c2488cfa24fd01ee88016a121479a2f86b394bd4. Its CONTRIBUTING.md requires synthetic people, companies, law firms, funds, products, addresses and matter facts and forbids real confidential client material. No real client data is present. Model: Qwen/Qwen3.5-9B (c202236235762e1c871ad0ccb60c8ee5ba337b9a); one tool call per turn; the system message is the collection prompt (the frozen evaluation prompt plus one sentence forbidding the model from writing its own <tool_result> block). Tasks were generated from specs of three tiers (mop-up, resample, relational) over the corpus; the 250 held-out benchmark tasks were decontaminated against by 8-gram overlap at generation time.

Schema

column type
rollout_id string session id, e.g. r5_0523_s1 (spec id + sample); recall episodes are p6_<rollout_id>_s<k>
mode string task family and tier, e.g. D5, R2, G1
family string D, R or G
task string the user instruction, verbatim
status string done (a <final> answer) or no_final
n_turns int32 assistant turns
n_tool_calls int32 tool calls executed
recall_episode_ids list recall episodes cut from this session
recall_judged_keep_ids list those that passed the honesty judge
final string the text inside the closing <final>...</final>, empty if none
transcript string the session rendered top-to-bottom: task, each assistant turn (thinking + call), each tool result (read this one)
events list<{step, role, content}> the structured session, lossless: user (task) then assistant / tool pairs per step

Source

pools/rollouts-v4/*.jsonl and pools/merged-v4/manifest.jsonl from the WM-RL v4 run directory on TACC Vista, exported 2026-09-15. Ten zstd parquet shards, 5,000 sessions each.

Downloads last month
17