ginigen/Korean-Hallucination-Bench
收藏资源简介:
Korean Hallucination Benchmark(한국어 환각 진단 벤치마크)是一个用于评估韩语大语言模型幻觉抵抗性的四选一多项选择题基准测试。它涵盖五个专业领域:法律、专利、行政、医疗和金融,并包含10种韩语特定幻觉类型,如数值/日期错误、条款扭曲、无根据推理、汉字词/新词混淆、过度泛化、多轮上下文崩溃、敬语/非敬语反转、来源伪造、术语扭曲和事实伪造。每个问题由一个事实正确答案和三个看似合理、足以欺骗专家的幻觉错误答案组成。数据集总计10,167个问题,通过Darwin-398B-JGOS生成问题与选项,并使用Claude进行答案验证和校正以确保准确性,同时自动丢弃答案不明确或存在多个正确答案的问题。数据以JSON格式存储,包括领域、幻觉类型、问题、选项、答案索引、答案文本和解释。评估方法基于greedy decoding,测量模型在避免选择幻觉选项的情况下选择正确答案的准确率。该数据集主要用于GenizenAI的LLM X-RAY幻觉诊断模块,以诊断韩语LLM在不同领域的幻觉漏洞。许可证为Apache 2.0,由GenizenAI在NIPA的先进GPU利用支持项目下开发。
The Korean Hallucination Benchmark (한국어 환각 진단 벤치마크) is a multiple-choice benchmark with four options designed to evaluate the hallucination resistance of Korean large language models. It covers five specialized domains: legal, patent, administrative, medical, and finance, and includes 10 Korean-specific hallucination types, such as numerical/date errors, clause distortion, unfounded reasoning, hanja/neologism confusion, overgeneralization, multi-turn context collapse, honorific/casual speech reversal, source fabrication, term distortion, and fact fabrication. Each question consists of one factual correct answer and three plausible hallucination incorrect answers that are convincing enough to deceive experts. The dataset contains a total of 10,167 questions, generated using Darwin-398B-JGOS for questions and options, with answer validation and correction performed by Claude to ensure accuracy, and automatic discarding of questions with ambiguous or multiple correct answers. Data is stored in JSON format, including domain, hallucination type, question, options, answer index, answer text, and explanation. Evaluation is based on greedy decoding, measuring the models accuracy in selecting the correct answer while avoiding hallucination options. The dataset is primarily used for GenizenAIs LLM X-RAY hallucination diagnosis module to identify hallucination vulnerabilities in Korean LLMs across different domains. It is licensed under Apache 2.0 and developed by GenizenAI with support from NIPAs Advanced GPU Utilization Support Project.





