遇见数据集

CPT 评测集 RAG 冲突 (DRAGged-into-Conflicts)

收藏
魔搭社区2026-06-04 更新2026-07-15 收录
官方服务:

资源简介:

# CPT-Eval-RAGConflicts [DRAGged-into-Conflicts](https://arxiv.org/abs/2406.16289) split used as the zero-shot RAG transfer benchmark in [Cognitive Pairwise Training (CPT)](https://github.com/Tsinghua-dhy/CPT) (paper Table 4). The CPT model is evaluated **without any RAG-specific training** — the abstention boundary learned during CPT mid-training is expected to transfer. ## Files - `conflicts.jsonl` — full split, ~46 MB ## Schema ```json { "question": "...", "documents": [...], "correct_answer": "...", "conflict_type": "No conflict | Conflict due to misinformation | Conflict due to outdated information | Complementary information | Conflicting opinions and research outcomes" } ``` ## Use ```bash modelscope download --dataset Tsinghuadhy/CPT-Eval-RAGConflicts --local_dir CPT/eval/dataset/conflicts cd CPT/eval/rag bash eval_rag_conflicts.sh Tsinghuadhy/CPT-RL-Qwen3-4B ``` ## Source Original *DRAGged-into-Conflicts*: <https://github.com/AviAvni/DRAGged-into-Conflicts>

提供机构:
maas
创建时间:
2026-06-01
二维码
社区交流群
二维码
科研交流群
商业服务