---
title: CPT-Pairs-90K
canonical_url: "https://www.modelscope.cn/datasets/Tsinghuadhy/CPT-Pairs-90K"
md_url: "https://www.modelscope.cn/datasets/Tsinghuadhy/CPT-Pairs-90K.md"
repository: Tsinghuadhy/CPT-Pairs-90K
chinese_name: "CPT 90K 推理路径配对数据（未判定）"
last_updated: 2026-06-01
license: MIT
storage_size: "555 MB"
downloads: 12
stars: 0
---

# CPT-Pairs-90K

> CPT-Pairs-90K - Tsinghuadhy 在 ModelScope 开源的数据集。Reasoning-trace pairs used by CPT (before judging).

Tsinghuadhy/CPT-Pairs-90K 是 ModelScope 魔搭社区上的数据集，存储大小 555 MB，采用 MIT 许可。

- **Repository**: Tsinghuadhy/CPT-Pairs-90K
- **License**: MIT
- **Storage size**: 555 MB
- **Downloads**: 12
- **Stars**: 0
- **Last updated**: 2026-06-01

Source: https://www.modelscope.cn/datasets/Tsinghuadhy/CPT-Pairs-90K

---

# 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