---
title: CPT-Pairs-Judged-77K
canonical_url: "https://www.modelscope.cn/datasets/Tsinghuadhy/CPT-Pairs-Judged-77K"
md_url: "https://www.modelscope.cn/datasets/Tsinghuadhy/CPT-Pairs-Judged-77K.md"
repository: Tsinghuadhy/CPT-Pairs-Judged-77K
chinese_name: "CPT 77K 推理路径配对（235B 共识判定）"
last_updated: 2026-05-31
license: MIT
storage_size: "895 MB"
downloads: 12
stars: 0
---

# CPT-Pairs-Judged-77K

> CPT-Pairs-Judged-77K - Tsinghuadhy 在 ModelScope 开源的数据集。CPT-Pairs-Judged-77K

Tsinghuadhy/CPT-Pairs-Judged-77K 是 ModelScope 魔搭社区上的数据集，存储大小 895 MB，采用 MIT 许可。

- **Repository**: Tsinghuadhy/CPT-Pairs-Judged-77K
- **License**: MIT
- **Storage size**: 895 MB
- **Downloads**: 12
- **Stars**: 0
- **Last updated**: 2026-05-31

Source: https://www.modelscope.cn/datasets/Tsinghuadhy/CPT-Pairs-Judged-77K

---

# 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