{"data":{"kind":"file","path":"README.md","version_id":"vbysdmpjo3zyb13826p7xuy7","entry":{"name":"README.md","path":"README.md","is_directory":false,"size":1337,"modified_at":"2026-11-06T19:52:25.400000","content_hash":"3ca9179e7eeb6935f3dc7d2a858613b11c26295a1240afbff676c5e378a4eb9b"},"entries":[],"content":"# fda-gym\n\nPublic [FDA-Gym](https://github.com/juanpablozarza/fda-gym) eval for Prime Intellect's Environments Hub. An agent answers regulatory questions with five tools over a frozen openFDA snapshot, then `submit_answer`. The grader scores the answer and the trace.\n\nThe published sheet is 93 tasks, `FG-9101`–`FG-9193` (50 `compare` and 43 `sponsor_hop`). Grader v5 is `fdagym.rewards.score()`. Freeze `2026-10-15` covers `drug/label`, `drug/event`, and `drug/drugsfda`. The prompt tells the model it may make 8 tool calls besides `submit_answer`. Ids outside the public range are refused.\n\nSource: [juanpablozarza/fda-gym](https://github.com/juanpablozarza/fda-gym). See [verifiers Discussion #3071](https://github.com/PrimeIntellect-ai/verifiers/discussions/3071).\n\n## Install\n\n```bash\nprime env install juanpablozarza/fda-gym\nuv run vf-eval fda-gym --dry-run --no-push\nuv run vf-eval fda-gym --select.include.ids FG-9101\n```\n\nThe environment wheel includes `fda_gym/data/tasks/public.jsonl` and `fda_gym/data/snapshots/2026-10-15`. A git checkout is not required. Set `FDAGYM_SNAPSHOT_DIR` or `FDAGYM_PUBLIC_JSONL` to override those files.\n\n## From a clone\n\n```bash\nuv sync --extra prime\nuv run vf-eval fdagym --dry-run --no-push\nuv run vf-eval @ configs/prime.toml --dry-run --no-push\n```\n\nA clone uses the repo's `data/` tree.\n","encoding":"utf-8","truncated":false,"total_bytes":1337},"status":null}