FLAMES
收藏资源简介:
FLAMES是一个针对中文大型语言模型(LLMs)的价值对齐基准数据集,由上海人工智能实验室和复旦大学联合创建。该数据集包含2,251个精心设计的手工对抗性提示,旨在评估LLMs与人类价值观的深度对齐,特别强调无害性和道德维度,如和谐等中国特定价值观。数据集通过复杂场景和破解方法设计对抗性提示,收集了17个主流LLMs的响应,并进行了严格标注以进行详细评估。FLAMES不仅揭示了现有LLMs在安全性和公平性方面的相对较差表现,还开发了一个轻量级的专用评分器,能够跨多个维度高效评估新模型,为当代LLMs的进一步对齐设定了新的挑战。
FLAMES is a value alignment benchmark dataset for Chinese large language models (LLMs), jointly created by Shanghai AI Laboratory and Fudan University. This dataset contains 2,251 meticulously designed handcrafted adversarial prompts, which are intended to evaluate the deep alignment between LLMs and human values, with a particular emphasis on harmlessness and ethical dimensions including China-specific values such as harmony. The adversarial prompts are designed based on complex scenarios and jailbreaking methods, and responses from 17 mainstream LLMs have been collected and rigorously annotated for detailed evaluation. FLAMES not only reveals the relatively poor performance of existing LLMs in terms of safety and fairness, but also develops a lightweight dedicated scorer that can efficiently evaluate new models across multiple dimensions, setting new challenges for further alignment of contemporary LLMs.

- 1Flames: Benchmarking Value Alignment of LLMs in Chinese上海人工智能实验室 复旦大学 · 2024年



