C-Plus Values
收藏资源简介:
C-Plus Values是一个全新的中文人价值观与大型模型对齐的评估基准,通过多轮对话和故事场景模拟现实世界情况,评估大型语言模型的责任感。该数据集由两部分组成:基于多轮对话的责任评估和基于故事场景的责任评估。它不仅要求模型避免生成有害内容,还强调共情和一定程度的人文关怀。数据集通过将CVALUES专家提出的问题转化为负向观点,并利用GPT-4 API生成多轮对话和故事格式的问题来构建。
C-Plus Values is a novel Chinese benchmark for evaluating the alignment between human values and large language models (LLMs). It simulates real-world scenarios through multi-turn dialogues and story-based contexts to assess the sense of responsibility of large language models. This dataset consists of two core components: responsibility evaluation based on multi-turn dialogues, and responsibility evaluation based on story scenarios. It not only mandates models to refrain from generating harmful content, but also places emphasis on empathy and a certain level of humanistic care. The dataset is constructed by converting questions proposed by CVALUES experts into negative viewpoints, and generating multi-turn dialogue and story-formatted questions via the GPT-4 API.

- 1Beyond Single-Sentence Prompts: Upgrading Value Alignment Benchmarks with Dialogues and Stories天津大学 · 2025年



