CPT 70K 监督微调数据

# 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.