CPT 评测集 数学基准

# CPT-Eval-Math Math reasoning evaluation suite used in [Cognitive Pairwise Training (CPT)](https://github.com/Tsinghua-dhy/CPT) (paper §4.2). 7 benchmarks used in the paper's main results: | Folder | Benchmark | # items | |----------------|------------------------|---------| | `math500/` | MATH-500 | 500 | | `aime24/` | AIME 2024 | 30 | | `aime25/` | AIME 2025 | 30 | | `amc22/` | AMC 2022 | 83 | | `amc23/` | AMC 2023 | 83 | | `minervamath/` | Minerva-Math | 272 | | `olympiad/` | OlympiadBench (math) | 675 | Each folder ships a single `test.jsonl` with one example per line. ## Schema ```json { "question": "...", "answer": "..." } ``` ## Use Pair this with [`eval/math/eval_math.sh`](https://github.com/Tsinghua-dhy/CPT) in the main repo: ```bash modelscope download --dataset Tsinghuadhy/CPT-Eval-Math --local_dir CPT/eval/dataset cd CPT/eval/math bash eval_math.sh Tsinghuadhy/CPT-RL-Qwen3-14B ```