遇见数据集

CPT 90K 推理路径配对数据(未判定)

收藏
魔搭社区2026-07-03 更新2026-07-15 收录
官方服务:

资源简介:

# CPT-Pairs-90K Reasoning-trace pairs used by [CPT](https://github.com/Tsinghua-dhy/CPT) (before judging). 90,970 pairs constructed over [CPT-Source-8556](https://www.modelscope.cn/datasets/Tsinghuadhy/CPT-Source-8556) via three pair-construction strategies: | Strategy | Count | |---|---| | Task 1 — Intra-Model | 39,042 | | Task 2 — Inter-Model | 39,042 | | Task 3 — Small-correct vs Large-wrong | 12,886 | Rollouts are produced by Qwen3-{4B, 8B, 14B}-Base. ## Format JSONL. Key fields: `question`, `ground_truth`, `difficulty`, `pair_type`, `task_type`, `path_a`, `path_b`. Each `path_*` contains: `model_name`, `model_size`, `rollout_idx`, `pred_ans`, `gen_text`, `is_correct_gpt`. ## Files - `data.jsonl` — 90,970 lines, ~556 MB

提供机构:
maas
创建时间:
2026-06-01
二维码
社区交流群
二维码
科研交流群
商业服务