kaushik-harsh-99/Uncensored-SFT-v2

这是一个名为高质量未审查指令数据集V2的英文指令数据集，主要用于指令跟随和模型对齐研究。数据集通过语义去重技术（相似度阈值0.90）对V1版本进行优化，移除语义相似的提示，以提高多样性和信息密度，同时保留有用的变体。它包含输入/输出对的JSONL格式数据，适用于监督微调（SFT）、指令调优、QLoRA、对齐实验等，旨在减少重复监督信号，增强训练效率，并帮助模型恢复指令跟随能力、减少过度拒绝行为。数据集侧重于提升提示覆盖范围，优先考虑多样性而非数据量大小。This is an English instruction dataset named High Quality Uncensored Instruction Dataset V2, primarily designed for instruction following and model alignment research. The dataset optimizes the V1 version through semantic deduplication (similarity threshold 0.90), removing semantically similar prompts to enhance diversity and information density while preserving useful variations. It contains JSONL-formatted input/output pairs and is suitable for supervised fine-tuning (SFT), instruction tuning, QLoRA, alignment experiments, etc. It aims to reduce repeated supervision signals, improve training efficiency, and help models recover instruction-following capabilities and reduce over-refusal behavior. The dataset focuses on increasing prompt coverage, prioritizing diversity over dataset size.