SAGED：面向语言模型的可定制公平性校准的整体性偏见基准测试流水线
SAGED: A Holistic Bias-Benchmarking Pipeline for Language Models with Customisable Fairness Calibration
AI总结：
SAGED是首个整体性偏见基准测试流水线，通过五阶段流程及反事实分支等技术缓解评估偏差，揭示了8b级模型在G20国家情感分析及美国总统角色扮演中的显著偏见与放大效应。
摘要：
开发无偏见的大型语言模型被广泛认为至关重要，但现有的基准测试由于范围有限、数据污染以及缺乏公平性基线，在检测偏见方面存在不足。SAGED(bias)是首个解决这些问题的整体性基准测试流水线。该流水线包含五个核心阶段：抓取材料、组装基准测试、生成回复、提取数值特征以及使用差异指标进行诊断。SAGED 包含用于衡量最大差异的指标（如影响比率）和偏见集中度的指标（如 Max Z-scores）。注意到指标工具偏见和提示词中的上下文偏见可能会扭曲评估结果，SAGED 实施了反事实分支和基线校准来进行缓解。作为演示，我们在 G20 国家上使用 SAGED 测试了包括 Gemma2、Llama3.1、Mistral 和 Qwen2 在内的流行的 8b 级模型。通过情感分析，我们发现尽管 Mistral 和 Qwen2 比起 Gemma2 和 Llama3.1 表现出较低的最大差异和较高的偏见集中度，但所有模型对俄罗斯等国家都有明显的偏见，并且（除 Qwen2 外）对中国也有偏见。通过进一步实验让模型扮演美国总统，我们看到偏见被放大并朝异质方向转移。此外，我们看到 Qwen2 和 Mistral 没有参与角色扮演，而 Llama3.1 和 Gemma2 扮演 Trump 的强度明显高于 Biden 和 Harris，表明这些模型中存在角色扮演性能偏见。
英文摘要：
The development of unbiased large language models is widely recognized as crucial, yet existing benchmarks fall short in detecting biases due to limited scope, contamination, and lack of a fairness baseline. SAGED(bias) is the first holistic benchmarking pipeline to address these problems. The pipeline encompasses five core stages: scraping materials, assembling benchmarks, generating responses, extracting numeric features, and diagnosing with disparity metrics. SAGED includes metrics for max disparity, such as impact ratio, and bias concentration, such as Max Z-scores. Noticing that metric tool bias and contextual bias in prompts can distort evaluation, SAGED implements counterfactual branching and baseline calibration for mitigation. For demonstration, we use SAGED on G20 Countries with popular 8b-level models including Gemma2, Llama3.1, Mistral, and Qwen2. With sentiment analysis, we find that while Mistral and Qwen2 show lower max disparity and higher bias concentration than Gemma2 and Llama3.1, all models are notably biased against countries like Russia and (except for Qwen2) China. With further experiments to have models role-playing U.S. presidents, we see bias amplifies and shifts in heterogeneous directions. Moreover, we see Qwen2 and Mistral not engage in role-playing, while Llama3.1 and Gemma2 role-play Trump notably more intensively than Biden and Harris, indicating role-playing performance bias in these models.
机构、总结、中文摘要均由AI分析而来，可能存在错误，请以论文正文为准。
