SOS-BENCH
收藏资源简介:
SOS-BENCH是由Arthur AI、NYU和Columbia University共同创建的,用于评估大型语言模型(LLM)对齐性能的标准化、可重复的元基准。该数据集整合了19个现有的世界知识、指令遵循和安全基准,旨在提供一个全面的模型性能视图。数据集中的每个问题都包含真实的答案,并通过标准化准确率的平均值来报告聚合结果。SOS-BENCH的应用领域主要集中在大型语言模型的对齐研究,旨在解决模型在安全性、世界知识和指令遵循方面的具体问题。
SOS-BENCH is a standardized, reproducible meta-benchmark co-created by Arthur AI, New York University (NYU), and Columbia University for evaluating the alignment performance of Large Language Models (LLMs). This dataset integrates 19 existing world knowledge, instruction-following, and safety benchmarks, aiming to provide a comprehensive view of model performance. Each question in the dataset includes a ground-truth answer, and aggregated results are reported via the average of standardized accuracy scores. The primary application scope of SOS-BENCH focuses on alignment research for large language models, with the goal of addressing specific issues related to model safety, world knowledge, and instruction following.

- 1Style over Substance: Failure Modes of LLM Judges in Alignment BenchmarkingArthur AI, NYU, Columbia University · 2024年



