heretic-org/Semantic-Harmful
收藏资源简介:
该数据集包含来自两个源数据集(mlabonne/harmful_behaviors和mlabonne/harmless_alpaca)的提示之间的一对一语义匹配,其中一个提示是有害的,另一个是无害的。通过语义对齐,这些匹配对旨在减少噪声,以便更精确地比较有害和无害提示,用于安全研究、拒绝分析、激活差异研究等。数据集构建方法包括加载源数据、使用嵌入模型生成文本嵌入、归一化、计算语义相似度、匹配和阈值过滤。数据集包含元数据,如匹配对、相似度分数和原始索引,文件格式为CSV、JSON或TXT。
This dataset contains one-to-one semantic matches between prompts from two source datasets: mlabonne/harmful_behaviors and mlabonne/harmless_alpaca, where one prompt is harmful and the other is harmless. The goal is to align prompts that are semantically closest to create a more controlled comparison set for safety research, refusal analysis, and activation-difference studies. It reduces noise by pairing harmful and harmless prompts that are semantically close, focusing the comparison. The construction pipeline involves loading source datasets, creating text embeddings with an embedding model, normalizing embeddings, computing semantic similarity, matching prompts, and threshold filtering. The dataset includes metadata such as matched pairs, similarity scores, and original indices, available in CSV, JSON, or TXT formats.




