---
title: CPT-SFT-70K
canonical_url: "https://www.modelscope.cn/datasets/Tsinghuadhy/CPT-SFT-70K"
md_url: "https://www.modelscope.cn/datasets/Tsinghuadhy/CPT-SFT-70K.md"
repository: Tsinghuadhy/CPT-SFT-70K
chinese_name: "CPT 70K 监督微调数据"
last_updated: 2026-05-31
license: MIT
storage_size: "287 MB"
downloads: 10
stars: 0
---

# CPT-SFT-70K

> CPT-SFT-70K - Tsinghuadhy 在 ModelScope 开源的数据集。Pairwise mid-training data for CPT.

Tsinghuadhy/CPT-SFT-70K 是 ModelScope 魔搭社区上的数据集，存储大小 287 MB，采用 MIT 许可。

- **Repository**: Tsinghuadhy/CPT-SFT-70K
- **License**: MIT
- **Storage size**: 287 MB
- **Downloads**: 10
- **Stars**: 0
- **Last updated**: 2026-05-31

Source: https://www.modelscope.cn/datasets/Tsinghuadhy/CPT-SFT-70K

---

# CPT-SFT-70K

Pairwise mid-training data for [CPT](https://github.com/Tsinghua-dhy/CPT).

Sampled 10,000 consensus pairs from [CPT-Pairs-Judged-77K](https://www.modelscope.cn/datasets/Tsinghuadhy/CPT-Pairs-Judged-77K) and expanded to 70,352 SFT samples by extracting *all* highest-confidence judgments per pair.

## Format

Parquet. Each sample has `prompt` and `response`:

- `prompt` — wrapped in the same `User: ... Assistant:` template as math reasoning, asking the model to compare two reasoning paths and output `\boxed{Path A is better}` (or B / equally good / equally bad).
- `response` — analysis text + final judgment in `\boxed{}`.

The Path A/B order has been randomly swapped at construction time to remove positional bias.

## Files

- `train.parquet` — 70,352 samples, ~285 MB
- `test.parquet`  — held-out, ~2 MB

## Usage

Plug-and-play with [verl](https://github.com/volcengine/verl) FSDP SFT trainer. See [CPT/train](https://github.com/Tsinghua-dhy/CPT) for example scripts.