Adds Qwen3-8B **thinking-enabled** teacher histories for 400 existing NL2Bash V2 tasks: **1,600 graded attempts, 900 successful traces, 300 tasks solved within four attempts**.
This is an additive model-specific shard with the same 14 columns/Arrow types as `coder.parquet`, plus a manifest and short notes. The dataset README gets only a link and a clearly separated subset summary. Existing Coder traces, statistics, selection evidence and V3 work are unchanged.
Validation:
- Exactly 400 existing task IDs × 4 unique attempts; all 400 content hashes match the current public Coder metadata.
- All 1,600 original ATIF-v1.8 documents schema-validate, and trajectory text is unchanged.
- Every raw reward agrees with the original trial result and `verifier/reward.txt`; 48 timeouts and all 700 failures are retained.
- Of those timeouts, 46 have verifier reward 0 and 2 have reward 1. Raw grades are preserved; 898 positive attempts have no recorded exception for consumers requiring normal termin…
Inputs are packaged under setup_files/inputs, so setup never contacts Hugging Face. task.toml still declared HF_TOKEN = "${HF_TOKEN}", which Harbor requires to exist on the host. All 5,000 tasks: only that line removed (all other archive bytes identical); content_sha256 updated.
Fills 15,360 result records and 14,125 conversations from the original stage-6 evidence of swelego-recovery-20261015-c25g. The 1,331 unknown outcomes are recorded terminal errors and now score zero; kept tasks unchanged at 1,245.
Add endless_v3: Qwen3.8-27B pass@2 on v2 exclusions, 2,079 kept
Qwen3.8-27B pass@2 on the 186 tasks excluded from endless_v2: 121 newly solved, 2,079 kept, 65 excluded. Adds traces, trace statistics and contracts. endless_v1 README: note that the repository-leak check did not search /home and a rerun is pending.
Keep the task and archive Parquets, README, patcher, and cached images. Place the task-to-image index inside cached_images/ and remove the separate patch-provenance manifest and archival evidence files.
Publishes all 21,440 recorded Coder pass@4 conversations from 5,360 evaluated tasks. Keeps 704 tasks solved at least once, preserves 4,656 unsuccessful tasks and 95 prior exclusions in archive, and records 565 pending tasks separately. Includes frozen provenance, settings and output-cap caveat, reproducible selection, and measured trajectory statistics. Task payloads are unchanged. This partial teacher release does not claim complete current-profile validation.
Add1258tasks with a verifier-successful trace whose recorded steps have output<8192,prompt+output<=25000 and zero summaries. Also add1379-task summary-allowed view. Preserve original tasks.parquet and all traces. Publish all-attempt measurements,witness IDs,selector and length distributions; label this retrospective filtering rather than a capped rerun.
Document intended and actual output-limit evidence in all v2/v3 READMEs
Keep all 2,277 v2 tasks and add 216 newly solved tasks. Preserve historical traces and publish 2,024 new saved conversations. Mark 47 tasks and 94 attempts pending; report new evaluation statistics separately.
Include SWE LEGO rerun: 1245 kept tasks, 96 new traces and refreshed statistics
Include the final 736 Thinking outcomes. Keep 2,246 tasks with at least one success, preserve raw null verifier rewards and the existing terminal-error scoring policy, and synchronize task columns, selection, combined traces, statistics and evaluation plots.
Preserve all 65 v2 tasks and add 406 newly solved tasks. Include all 16,496 old and new attempts, per-task success counts, success distributions, trajectory statistics, plots, and reproducible task selection.
Align all v1 dataset READMEs with the agreed format
Standardize per-stage explanations, image counts, per-task resource tables and timing coverage across 11 v1 dataset cards. Resource distributions come from the published task configurations. Preserve historical validation gaps and mark unsupported measurements as Missing. No task payloads or archive membership changed.
Document verified original image counts for MiMo and SETA
Restore the compact stats-first README and the standard README, patcher, tasks and archive layout. Embed the image-digest mapping in patch.py. Audit evidence remains accessible at the linked original publication revision. The self-contained patcher reproduces both task and archive Parquets byte-for-byte; 2,509 retained tasks and 1,189 archives are unchanged.
