CPT 题源 OpenMath 33K（Pool 3 题源审计后保留）

# CPT-Source-OpenMath-33K Quality-audited subset of [NVIDIA OpenMathReasoning](https://huggingface.co/datasets/nvidia/OpenMathReasoning), used to build **Pool 3** of the CPT 2× corpus (paper §B.4.1 / Appendix G.2). The audit uses Qwen3-235B-A22B-Instruct-2507 (`J-OPENMATH-AUDIT`) with pessimistic 4-sample voting to keep only items whose ground truth is verifiable and whose statement is unambiguous. Starting from **38,272** non-MCQ problems, **33,801** are retained (88.3%); the remaining 4,471 are removed (mainly: GT is descriptive natural language, ambiguous statements, missing figures, proof-style prompts that leak the target). The 8K difficulty-balanced **Pool 3** used in CPT 2× is sampled from `kept.jsonl`. ## Files | File | Lines | Description | |---|---:|---| | `kept.jsonl` | 33,801 | Audited & retained problems (this is what Pool 3 samples from) | | `removed.jsonl` | 4,471 | Removed problems with per-rollout audit decisions | | `audit_stats.json` | — | Per-reason removal breakdown (matches paper Table 40) | ## Schema ```jsonc { "question": "...", "answer": "...", "source": "openmath", "subject": "aops_c6_high_school_olympiads", "difficulty": "easy | medium | hard | very_hard", "accuracy": 0.4375, // Qwen3-72B-Instruct accuracy (20 rollouts) "correct_count": 9, "rollout_num": 20, "accuracy_72b_original": 0.6875, "filter_results": { "final_keep": true, "all_keep_decisions": [true, true, true, true], "all_categories": ["C","C","C","C"], // A=remove, B=borderline, C=keep "all_confidences": [0.98, 0.95, 0.98, 0.95], "all_reasons": ["..."], "sampling_method": "pessimistic" } } ``` See [Cognitive Pairwise Training](https://github.com/Tsinghua-dhy/CPT) and the `J-OPENMATH-AUDIT` prompt in paper Appendix.