Verified Dataset

Turn passing attempts from difficult, tool-free tasks into conversational SFT JSONL. The environment supplies repeated evidence; learning-zone selection avoids overweighting already saturated rows.

Files

• basic.py — verify, select the learning zone, and export passing attempts.
• curate_learning_zone.py — make the strict partial-rate curation rule explicit.
• export_manifest.py — pair the dataset and provenance sidecar with a compact manifest.

When to use

Use this after a task set produces real disagreement and the passing assistant responses are suitable supervision. learning_zone() selects task rows; passing-only export filters the attempts within those rows.

Tool traces from _26_multi_step_tools/ are excluded because the portable exporter is text-only. The resulting dataset can be checked in _28_ci_gating/ or handed to a trainer separately.

Run

python cookbook/environments/_27_verified_dataset/basic.py
python cookbook/environments/_27_verified_dataset/curate_learning_zone.py
python cookbook/environments/_27_verified_dataset/export_manifest.py

Requires OPENAI_API_KEY. Export creates a dataset and provenance artifacts; it does not train or update the running agent.