forseasons/dataagent-dabench-altimate-noctx-sft-410

DataAgent DAB Altimate NoCtx SFT 410

This dataset contains 410 validated SFT trajectories generated by an Altimate-centered DataAgent pipeline. The paired RL release is forseasons/dataagent-dabench-altimate-noctx-rl-500.

Contents

• —train_sft.jsonl: final SFT rows, one selected trajectory per task.
• —train_sft.parquet: parquet version of the same final rows.
• —summary.json: merge/filter summary.
• —dataset_package_manifest.json: row counts, statistics, checksums.
• —audit/: GLM judge packets/results and raw trajectory audit files.

Format

Rows use a ChatML-style message list with assistant XML tool calls in qwen3_xml style. Reasoning fields are process-faithful action rationales / trajectory rationales, not a claim of full private teacher chain-of-thought.

Statistics

• —Rows: 410
• —Unique tasks: 410
• —Tool format: qwen3_xml
• —Mean turns: 27.09
• —Mean assistant XML tool calls: 12.43
• —Mean GLM judge score: 0.888

Filtering

The trajectories were selected with Altimate no-context expert rollout, local contract checks, GLM-5.2 trajectory judge, self-consistency selection, and paired RL validator checks. All final rows passed the paired RL validate.py artifacts during merge.