{"data":{"kind":"file","path":"README.md","version_id":"dvjcrgzs3pgvu0k841j1d27k","entry":{"name":"README.md","path":"README.md","is_directory":false,"size":1640,"modified_at":"2026-04-29T18:57:41.462000","content_hash":"427c5f5f9597637193456633df165efd91d248f92c8f3ae1efa22cb026e59604"},"entries":[],"content":"# countdown-plain\n\nCountdown arithmetic environment adapted for older/plain chat models.\n\nThis is a sibling of `sami/countdown` with:\n\n- no system message by default\n- plain instructions prepended to the user question\n- no XML or tag format requirement\n- a permissive verifier that extracts the last arithmetic expression from prose\n- easier default puzzles with 4 numbers instead of classic 6-number Countdown\n\nExample accepted completions:\n\n```text\nI can add 8 and 7 to get 15. The final expression is 8 + 7.\nFirst multiply 5 by 4, then add 3. Final answer: 5 * 4 + 3.\n```\n\nThe verifier checks that the expression:\n\n- uses only the provided numbers\n- uses each number at most once\n- uses ordinary arithmetic operations\n- evaluates to the target for exact reward\n\nIt also includes a small closeness reward for valid expressions that miss the\ntarget.\n\n## Usage\n\n```python\nfrom verifiers import load_environment\n\nenv = load_environment(\"countdown-plain\")\n```\n\nWith the Talkie GRPO smoke script:\n\n```bash\npython examples/grpo_lewtun_hf_smoke.py \\\n  --vf-env sivit/countdown-plain \\\n  --vf-dataset-size 64\n```\n\n## Environment Arguments\n\n| Arg | Type | Default | Description |\n| --- | --- | --- | --- |\n| `num_examples` | int | `400` | Number of training examples |\n| `eval_examples` | int | `80` | Number of eval examples |\n| `seed` | int | `42` | Random seed |\n| `num_numbers` | int | `4` | Numbers available in each puzzle |\n| `num_large` | int | `1` | Count of large numbers sampled from 25, 50, 75, 100 |\n| `exact_weight` | float | `1.0` | Exact expression reward weight |\n| `closeness_weight` | float | `0.25` | Closeness reward weight |\n","encoding":"utf-8","truncated":false,"total_bytes":1640},"status":null}