{"data":{"kind":"file","path":"README.md","version_id":"ep1sbp5pwbeekfe8r3c8ef3z","entry":{"name":"README.md","path":"README.md","is_directory":false,"size":2399,"modified_at":"2026-06-03T18:48:17.722000","content_hash":"251e6816dbe48bb37e870214a1c033aba87932ddfed0af1f2b960da4aff55aa9"},"entries":[],"content":"# AIME-26\n\n<a href=\"https://github.com/PrimeIntellect-ai/research-environments/tree/main/environments/aime2026\">\n<img src=\"https://img.shields.io/badge/GitHub-181717?style=for-the-badge&logo=github&logoColor=white\" alt=\"Source Code\">\n</a>\n\n### Overview\n- **Environment ID**: `aime2026`\n- **Short description**: AIME 2026 problems (AIME I/II) evaluated single-turn with CoT and boxed numeric answers.\n- **Tags**: math, aime, 2026, single-turn, boxed-answer\n\n### Datasets\n- **Primary dataset(s)**: [MathArena/aime_2026](https://huggingface.co/datasets/MathArena/aime_2026) loaded directly via `load_dataset` and pinned to revision `10b4e45b7a503075d4da8a0d57916a4f06ce6bd2`\n- **Split sizes**: Defaults to split `train` (N=30)\n\n### Task\n- **Type**: single-turn\n- **Answer extraction**: `extract_boxed_answer` on the final assistant message\n- **Reward overview**: the `math_verify` reward compares the boxed answer to the target (single criterion, weight 1.0).\n\n### Quickstart\nRun an evaluation with default settings:\n\n```bash\nuv run vf-eval aime2026\n```\n\nConfigure model and sampling:\n\n```bash\nuv run vf-eval aime2026 \\\n  -m gpt-4.1-mini \\\n  -n 20 -r 3 -t 1024 -T 0.7\n```\n\nNotes:\n- Use `-a` / `--env-args` to pass environment-specific configuration as a JSON object.\n\n### Environment Arguments\nNone. The dataset, prompt, and grading settings are fixed for this benchmark.\n\n### Metrics\n| Metric | Meaning |\n| ------ | ------- |\n| `reward` | Weighted sum of rewards (here, `math_verify`) |\n| `math_verify` | 1.0 if the boxed answer matches the target under `math_verify`, else 0.0 |\n\n### Changelog\n\n### v0.2.0\n- Port to the Taskset/Harness environment interface.\n- Bump verifiers to `0.1.15.dev152`.\n\n### v0.1.3\n- Fix scoring bug: explicitly cast `answer` column to string type in dataset mapping to prevent HuggingFace Datasets from coercing `str(int(...))` back to `int64`, which caused `'int' object is not subscriptable` errors in the rubric\n\n### v0.1.2\n- Pin HuggingFace dataset loading to a fixed revision and set `trust_remote_code=False`\n\n### v0.1.1\n- Bump verifiers to v0.1.12.dev1: perf improvements to `MathRubric`; now uses `extract_boxed_answer` in strict mode — if no `\\boxed{}` answer is found the parsed answer is `\"\"` which always scores 0, preventing false positives where the model is rewarded for containing the correct answer anywhere in the response\n\n### v0.1.0\n- Initial release\n","encoding":"utf-8","truncated":false,"total_bytes":2399},"status":null}