CPT 77K 推理路径配对（235B 共识判定）

# CPT-Pairs-Judged-77K Self-consistency–judged reasoning-trace pairs used by [CPT](https://github.com/Tsinghua-dhy/CPT). Built on top of [CPT-Pairs-90K](https://www.modelscope.cn/datasets/Tsinghuadhy/CPT-Pairs-90K). Each pair is judged 8 times by Qwen3-235B-Instruct (4 dimensions: correctness, logical soundness, clarity, efficiency); only pairs reaching consensus (≥5/8 agreement) are kept. 77,657 pairs (consensus rate 85.37%). ## Format JSONL. Top-level fields are inherited from `CPT-Pairs-90K`. The `judge_results` dict is slimmed to: - `final_judgment` — one of `Path A is better` / `Path B is better` / `Both are equally good` / `Both are equally bad` - `consensus_count`, `total_rollouts` - `judgment_difficulty`, `judgment_metrics` - `analysis` — analysis from the highest-confidence round whose judgment matches `final_judgment` - `confidence` — `Very High` / `High` / `Medium` / `Low` - `confidence_analysis` ## Files - `data.jsonl` — 77,657 lines, ~800 MB