·
2 commits
to main
since this release
349 full LLM-played games of Balatro (Red Deck, White Stake) on the jackdaw simulator, harvested with scripts/llm_run.py; 200 wins, of which 189 replay bit-exact via scripts/demo_replay.py and form the training demo set. Per game: a human-readable move log and a structured per-move pairs_*.jsonl with the model's full reasoning. Models: gemini-3.7-flash (307 games / 181 wins), deepseek-v4-pro-0813 (36/18), plus a few pilot games (claude-sonnet-5, gemini-3.1-pro-preview, glm-5.2). MANIFEST.csv indexes every game; the archive's README documents format, verification, and how to rebuild the training pickles. License: CC0 1.0.