{"data":{"kind":"file","path":"README.md","version_id":"qvotpk1s8nixei8c6a64kz8g","entry":{"name":"README.md","path":"README.md","is_directory":false,"size":1228,"modified_at":"2026-08-03T03:23:19.688000","content_hash":"2b48198ffcb24dfeb122d23ad08e2ccf105b0e38f7d3c0b921231c07403bfbef"},"entries":[],"content":"# catan-v1\n\nA Settlers of Catan environment for training and evaluating language models.\n\nEach rollout is a full four-player game. The model sees a text description of the board and chooses from a numbered list of legal moves. Players only trade with the bank.\n\nOn Prime, one LLM plays against three bots. Run `CatanEnv` directly for games with multiple LLMs.\n\n## reward\n\nThe main reward is winning the game. A small normalized victory-point reward provides denser feedback:\n\n- `reward_win`: `1` for a win, otherwise `0`\n- `reward_vp`: `min(victory_points, 10) / 10`, weighted by `vp_coef`\n\nTruncated games keep their victory-point.\n\n## seeds\n\nSeeds `0–9999` are reserved for evals. Training must start at `10000` or above. The loader rejects ranges that cross the boundary.\n\n## hosted setup\n\nSet `PYTHONHASHSEED=0` once so action ordering stays deterministic:\n\n```bash\nprime env var create taziksh/catan-v1 \\\n  --name PYTHONHASHSEED \\\n  --value 0\n```\n\n```toml\n[[env]]\nid = \"taziksh/catan-v1\"\nargs = { seed_start = 10000, seats = \"agent,value_function,value_function,value_function\", invalid_retries = 1, vp_coef = 0.1, max_turns = 500 }\n```\n\nOther args: `num_seeds`, `trajectory_dir`, `system_prompt`, and `timeout_seconds`.\n","encoding":"utf-8","truncated":false,"total_bytes":1228},"status":null}