MarkrAI/k-overrefusal
收藏资源简介:
k-overrefusal是一个专门用于评估韩语大型语言模型(LLM)过度安全拒绝(over-refusal)或错误拒绝(false refusal)的基准测试数据集。该数据集旨在测量模型在无害但表面看似有害的请求上错误拒绝的行为,同时确保模型能正确拒绝真正有害的请求。数据集包含485个样本,分为两类:safe(无害但看似有害的请求,共185个)和unsafe(真正有害的请求,共300个)。safe样本进一步根据语言和上下文机制细分为多个类别,包括k_idioms(韩语惯用和文化表达)、legality_framing(合法框架)、privacy(隐私)、safe_contexts(安全上下文)、euphemism(委婉语)、harmless_analog(无害类比)、benign_purpose(良性目的)和other(其他)。unsafe样本包括contrast_k_idioms(与k_idioms对应的真正有害请求)和从多个公开有害基准(如HarmBench、StrongREJECT等)中匹配的harm_*类别。数据集的构建过程涉及从公开过度拒绝数据集中筛选困难样本,并使用多语言句子嵌入进行语义匹配,以确保无害和有害请求之间的对比性。评估方法基于模型对prompt_ko(韩语输入)的响应,使用K-SafeGuard作为评判工具,测量safe样本的合规性(compliance)和unsafe样本的拒绝率(refusal)。数据集主要用于韩语LLM和安全防护的评估与研究,包含实际有害请求,因此仅限于安全评估和红队测试目的使用。
k-overrefusal is a benchmark dataset designed to evaluate over-refusal or false refusal in Korean large language models (LLMs). It measures the tendency of models to incorrectly reject harmless but seemingly harmful requests while ensuring proper refusal of genuinely harmful requests. The dataset consists of 485 samples divided into two categories: safe (pseudo-harmful, harmless but appearing harmful, 185 samples) and unsafe (harmful, genuinely harmful, 300 samples). Safe samples are further categorized by linguistic and contextual mechanisms, including k_idioms (Korean idiomatic and cultural expressions), legality_framing, privacy, safe_contexts, euphemism, harmless_analog, benign_purpose, and other. Unsafe samples include contrast_k_idioms (genuinely harmful requests corresponding to k_idioms) and harm_* categories matched from multiple public harmful benchmarks (e.g., HarmBench, StrongREJECT). The dataset construction involves selecting hard samples from public over-refusal datasets and using multilingual sentence embeddings for semantic matching to create precise contrast pairs. Evaluation is based on model responses to prompt_ko (Korean input) using K-SafeGuard as a judge, measuring compliance for safe samples and refusal for unsafe samples. The dataset is intended for evaluating and researching Korean LLM safety and guardrails, containing actual harmful requests, and should be used only for safety assessment and red-teaming purposes.




