decision_feedback

Tool tools/llm/decision_feedback.py (program, Models). Save explicit decision answer corrections per OS user and export a local SFT dataset.

Run it with ./gameframe run decision_feedback ARGS or python3 tools/llm/decision_feedback.py. Source on GitHub.

Its own help

This is the help text in the tool itself, as ./gameframe tool decision_feedback prints it.

Save explicit decision answer corrections per OS user and export a local SFT dataset.

Usage:
  decision_feedback.py record REQUEST.json --question ID --answer KEY [--model ID]
  decision_feedback.py export --output NEW_DIRECTORY [--valid-fraction 0.1]

The record command stores one user-supplied corrected answer. It never records model
responses automatically. The default ledger is private to the current OS account under
its application-data directory; set GAMEFRAME_DECISION_DATA_DIR or pass --data-dir to choose
another profile. The state and question text are saved there, so record only material the
user wants kept.

The export command keeps the latest correction for each exact state/question/options prompt.
It writes chat-format JSONL files for local fine-tuning tools and a manifest. Fewer than ten
distinct states produce a train file without validation data. This tool prepares examples; it
does not update model weights, install a trainer, upload data, or start a paid training job.