{"data":{"kind":"file","path":"README.md","version_id":"j0aeb122pogwc80tyw6ovw9a","entry":{"name":"README.md","path":"README.md","is_directory":false,"size":1234,"modified_at":"2026-06-04T00:34:28.650000","content_hash":"33309ca7fb67080a71ef5b00a4eb58a682666f8d753870f83a75078a5b006886"},"entries":[],"content":"# Adaption Dakota QA Environment\n\nPrime Intellect / Verifiers single-turn RL environment for `HarleyCooper/adaption-dakota-english-qa`.\n\nThe dataset contains 2,445 English-language question-answer tasks derived from the 1890 Dakota-English Dictionary and remastered by Adaption Labs. This environment adapts the new schema:\n\n- `question`: concise user question\n- `answer`: authoritative target answer used for reward\n- `enhanced_prompt`: richer instruction/context prompt, used by default\n- `enhanced_completion`: retained as metadata only, not treated as ground truth\n- source metadata: `pair_id`, `source_files`, `source_pages`, `source_language`\n\nThe reward is deterministic and continuous so hosted RL has useful variance: normalized exact-match, Dakota term coverage, character F1 against the target answer, lexical keyword coverage, and a brevity/readability component.\n\n## Local smoke\n\n```bash\nuv pip install -e .\nuv run vf-eval adaption_dakota_qa -n 5 -r 1\n```\n\n## Hosted training starter\n\n```bash\nprime train configs/rl/laguna-smoke.toml --plain -y\n```\n\n`poolside/Laguna-XS.2` is the intended free model from the Prime example. If it is at capacity, `sprints/Llama-3.2-1B-Instruct` is also currently listed by Prime as free.\n","encoding":"utf-8","truncated":false,"total_bytes":1234},"status":null}