{"data":{"kind":"file","path":"README.md","version_id":"ozkd5cuk3rbdux5gmk4bchtu","entry":{"name":"README.md","path":"README.md","is_directory":false,"size":1523,"modified_at":"2026-06-27T01:36:21.497000","content_hash":"752e7bdf56009e088c216be0976e70f3eb0cf0e062cc581d723dec5e87b4f4f4"},"entries":[],"content":"# Tycoon Learning Environment\n\n![TycoonLE replay interface](https://raw.githubusercontent.com/vrtnis/tycoon-learning-environment/main/assets/tycoonLE.png)\n\nPrime/Verifiers environment for evaluating LLM agents on TycoonLE logistics planning.\n\nThe model receives a TycoonLE world state and a visible table of executable candidate actions. On each turn it must return JSON:\n\n```json\n{\"action_index\": 3, \"reason\": \"short reason\"}\n```\n\nThe environment parses the action, executes it in TycoonLE, returns the updated state, and scores the rollout using final TycoonLE score, JSON validity, and valid-action rate.\n\n## Local Smoke Eval\n\nInstall Prime CLI and authenticate separately:\n\n```powershell\npip install prime\nprime config set-api-key\n```\n\nRun a small eval after the environment is installed:\n\n```powershell\nprime eval run tycoon-learning-environment -p .\\environments -n 3 -r 1 -m openai/gpt-4.1-mini --max-tokens 128\n```\n\nFor local development without spending credits, import `load_environment()` and call the environment parser/scorer directly.\n\n## Environment Args\n\n- `split`: TycoonLE split, default `dev`\n- `families`: comma-separated family names, default all families\n- `seed_start`: first procedural seed, default `20000`\n- `num_examples`: dataset size, default `5`\n- `candidate_limit`: maximum visible candidates per turn, default `12`\n- `action_budget`: maximum executed TycoonLE actions per rollout, default `6`\n- `max_turns`: Prime/Verifiers turn limit; when provided, sets `action_budget` to `max_turns - 1`\n","encoding":"utf-8","truncated":false,"total_bytes":1523},"status":null}