CPT 训练数据 Abstention-SFT baseline
收藏官方服务：
资源简介：
# cpt-trainingdata-abstention-sft Abstention-aware SFT data used to warm-start the `Abs-RL` baseline in [Cognitive Pairwise Training (CPT)](https://github.com/Tsinghua-dhy/CPT) (paper §4.3). A ~5K abstention-aware RL mixture (information-insufficient + false-premise + answerable math items), followed by GRPO Math-RL to produce the Abs-RL checkpoints. Train script: [`train/sft/run_abs_sft_qwen3_4b.sh`](https://github.com/Tsinghua-dhy/CPT). ## Files - `train.parquet` — ~224 MB
提供机构：
maas创建时间：
2026-06-01
