Koalacrown/conflict_pairs

该数据集包含从UltraFeedback数据集中生成的冲突对。通过筛选高分歧、高质量的回答对，并使用本地LLM通过vLLM推断出可能产生每个回答的对比指令来创建。生成流程包括：1. 加载UltraFeedback（64k提示×4个回答）；2. 预筛选高分歧、高质量的回答对；3. 使用本地LLM从每个回答对推断对比指令；4. 解析并筛选真正的冲突。This dataset contains conflict-pairs generated from the UltraFeedback dataset. It was created by filtering for high-divergence, decent-quality response pairs and using a local LLM via vLLM to infer contrasting instructions that could have produced each response. The pipeline includes: 1. Load UltraFeedback (64k prompts × 4 responses each); 2. Pre-filter to high-divergence, decent-quality pairs; 3. Use a local LLM to infer contrasting instructions from each pair; 4. Parse and filter for genuine conflicts.