grpo_training_data.jsonl: prompt, thought, action, reward, and observed state trajectories.
rl_trajectories_with_cot.jsonl: environment-side trajectory log with chain-of-thought separated from action JSON.
logs/collection_summary.json: aggregate data-collection metrics.
logs/grpo_reward_100_fixed.jsonl: reward log from the 100-step GRPO run.
logs/train_grpo_100_fixed.log: training log for the completed fixed GRPO run.
gol-grpo-100-fixed/: trainer metadata/config. Large model weights are included only when uploaded with --include-model and present locally.

Collection Summary

Total episodes: unknown
Reward min: unknown
Reward max: unknown
Reward mean: unknown
Empty thoughts: unknown
Format failures: unknown

Caveats

The first 100-step GRPO run completed, but generated completions were clipped at the configured maximum length. Later local code changes tightened prompt formatting, added format penalties, and hardened reward logic against detector/entity-count exploits.

brysgo/gol-grpo-diversified-37679423

GoL RL Run Artifacts

Contents

Collection Summary

Caveats

brysgo/
gol-grpo-diversified-37679423