Datasets:

---

FWeindel

/

validated-tasks

Like 0

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 termination.
• All 900 successful traces contain reasoning, tool calls and observations. The 26 failed attempts without a structured tool call also remain in the evidence.
• Exported records are read back and checked; operational host paths/endpoints and credential metadata are sanitized in companion trial records.

Recipe: Qwen/Qwen3-8B@b968826d9c46dd6066d109eabc6255188de91218; thinking on; Harbor 0.24.0 / Terminus-2; Docker fixture staging; 32K context; explicit 8,192 output tokens per call; temperature 0.6, top_p 0.95, top_k 20, min_p 0. Exact settings and task-level coverage are in the manifest.

Interpretation: 400 tasks were selected from the 938 V2 tasks already solved by the historical 30B teacher. Historical 30B pass@4 on this subset is 100% by construction; 75% is only the observed 8B pass@4 on this selected subset. Docker, thinking and response-limit differences prevent a controlled model comparison.

Suggested review: inspect the short notes and manifest, then load the new shard. No existing release artifact needs regeneration to use these attempts. This V2 contribution is separate from the ongoing NL2Bash V3 PR.