{"data":{"kind":"file","path":"README.md","version_id":"rlosb2a65itmuba4usm3te22","entry":{"name":"README.md","path":"README.md","is_directory":false,"size":1248,"modified_at":"2026-07-01T15:47:13.320000","content_hash":"0ecb307cdfca3860d9d5ed808df04bea6caba0e3880f606d06ac4854eefb9093"},"entries":[],"content":"# chess-rlvr\n\nSingle-turn chess RLVR environment for Prime Environment Hub.\n\nThe environment uses the design discussed for chess move-quality training:\n\n- model input: FEN plus legal SAN moves\n- model output: the selected SAN move\n- reward: precomputed Stockfish negative regret for the selected move\n\nThe default dataset is `albertklorer/chess-stockfish-regret`.\n\nEach dataset row contains:\n\n```json\n{\n  \"id\": \"00008\",\n  \"fen\": \"r6k/pp2r2p/4Rp1Q/3p4/8/1N1P2R1/PqP2bPP/7K b - - 0 24\",\n  \"legal_moves\": \"{\\\"Bc5\\\": -0.178, \\\"Qd4\\\": 0.0}\"\n}\n```\n\n`legal_moves` is a JSON-encoded object mapping SAN moves to regret scores.\nThe best available legal move has reward `0.0`; worse moves are negative.\n\n## Prompt shape\n\n```text\nFEN:\n<fen>\n\nLegal SAN moves:\n<SAN>\n<SAN>\n...\n\nReturn only the best SAN move.\n```\n\n## Environment args\n\n| Arg | Type | Default | Description |\n| --- | --- | --- | --- |\n| `dataset` | str | `albertklorer/chess-stockfish-regret` | Hugging Face dataset id |\n| `split` | str | `train` | Dataset split |\n\n## Build note\n\nThis is a standard single-file `verifiers` environment, not an OpenEnv server.\nUse `prime env push` to build and publish the wheel. `prime env build` is only\nfor OpenEnv-backed environments with a `proj/` directory.\n","encoding":"utf-8","truncated":false,"total_bytes":1248},"status":null}