training/evidence/README.md

Training Evidence

Small, replayable evidence.json files for the four method comparisons used in the demo:

• sft/
• dpo/
• sft_dpo/
• grpo_rlvr/

These files store recorded CoS action routes, rewards, fallback usage, and terminal scores. They do not require adapter weights, so the GRPO+RLVR evidence can be used even when the exported run has no adapter_config.json.

The plots/ folder contains small committed PNGs for judges:

• terminal scores by method
• policy rewards by method
• expert-brief reward curve
• RL loss by method
• RL best-reward tracking by method
• RL chosen-action correctness by method

The rl_training_metrics/ folder contains real train_metrics.json exports from GRPO, PPO, and GRPO+RLVR runs. These are the source for the RL loss plots.

Generate full textual context reports with:

python3 training/scripts/kaggle_context_results_from_evidence.py --roots .