{"data":{"kind":"file","path":"README.md","version_id":"kq74btfe56jjdzj5pi0o3yz7","entry":{"name":"README.md","path":"README.md","is_directory":false,"size":1635,"modified_at":"2026-06-15T16:59:29.295000","content_hash":"28c65086730c34b30d3b8fcdbc9ee4c62a92f6207953ec6898973f052bc5dbee"},"entries":[],"content":"# meta-sparse-shaping\n\n`meta-sparse-shaping` is a deterministic Verifiers environment for comparing\nsparse terminal reward with shaped intermediate reward on the same prompts.\n\nEach example gives an integer list and a sequence of list operations. The model\nmust return exactly one JSON object inside a `<result>...</result>` tag:\n\n```text\n<result>{\"trace\": [[...], [...]], \"answer\": [...]}</result>\n```\n\n`trace` is the list after each operation. `answer` is the final list.\n\nThe environment exposes two reward modes:\n\n- `reward_mode=\"sparse\"`: format/schema credit is constant, but task reward\n  mostly comes from exact final-answer correctness.\n- `reward_mode=\"shaped\"`: the same format/schema credit is used, plus partial\n  credit for final-list element accuracy and exact intermediate trace steps.\n\nThis creates paired RL runs with identical data distributions and different\nreward surfaces.\n\n## Usage\n\n```python\nfrom verifiers import load_environment\n\nenv = load_environment(\n    \"meta-sparse-shaping\",\n    seed=20260615,\n    num_examples=128,\n    min_length=4,\n    max_length=7,\n    min_ops=3,\n    max_ops=5,\n    operation_set=\"core\",\n    reward_mode=\"shaped\",\n)\n```\n\n## Operation Sets\n\n`operation_set=\"core\"` uses:\n\n- `add`\n- `subtract`\n- `multiply`\n- `reverse`\n- `rotate_left`\n- `sort_asc`\n\n`operation_set=\"extended\"` also adds:\n\n- `sort_desc`\n- `keep_even`\n- `keep_greater_than`\n\n## Recommended First Probes\n\nStart with both reward modes on the same seed using Qwen 2B or Llama 1B. The\nmain comparison metrics are reward, final exact answer, answer element accuracy,\ntrace step accuracy, schema validity, and exact-one-result.\n","encoding":"utf-8","truncated":false,"total_bytes":1635},"status":null}