FLEX
收藏资源简介:
FLEX(Fairness Benchmark in LLM under Extreme Scenarios)是一个专为评估大型语言模型在极端情况下公平性的基准数据集。该数据集由韩国大学的研究团队创建,旨在通过在问题中加入可能引发偏见的极端情景,来评估模型在面对这些极端条件下的公平性和鲁棒性。FLEX基于已有的公平性基准数据集,如BBQ、CrowS-Pairs和StereoSet,通过添加能够最大化模型漏洞的极端情景提示,重构了这些问题,形成了一个能够挑战模型在极端情况下保持中立和避免有害内容的鲁棒性评估。
FLEX (Fairness Benchmark in LLM under Extreme Scenarios) is a benchmark dataset dedicated to evaluating the fairness of large language models (LLMs) in extreme scenarios. Developed by a research team from South Korean universities, this dataset aims to assess the fairness and robustness of models when faced with extreme, bias-inducing scenarios embedded within test questions. Building upon existing fairness benchmark datasets including BBQ, CrowS-Pairs, and StereoSet, FLEX reconstructs these original questions by adding extreme scenario prompts that maximize the exposure of model vulnerabilities, creating an evaluation framework that challenges models to maintain neutrality and avoid generating harmful content under extreme circumstances.




