MM-RLHF
收藏资源简介:
MM-RLHF数据集是由中国科学院自动化研究所构建的,包含12万细粒度、人工注释的偏好比较对。该数据集在规模、多样性、注释粒度和质量上都有所提升,涵盖了图像、视频理解和多模态安全三个领域。数据集的创建经过严格的管道,包括数据收集、数据筛选、模型响应生成和精细的人类注释等步骤。该数据集旨在推动多模态大型语言模型(MLLM)的偏好对齐研究,解决模型在实际应用中的 truthfulness、safety 和对人类偏好的对齐等问题。
The MM-RLHF dataset was constructed by the Institute of Automation, Chinese Academy of Sciences, containing 120,000 fine-grained, manually annotated preference comparison pairs. This dataset features improvements in scale, diversity, annotation granularity and quality, covering three domains: image understanding, video understanding and multimodal safety. The development of this dataset follows a rigorous pipeline, including steps such as data collection, data filtering, model response generation and fine-grained human annotation. This dataset aims to promote research on preference alignment for Multimodal Large Language Models (MLLMs), and address issues including model truthfulness, safety and alignment with human preferences in real-world applications.

- 1MM-RLHF: The Next Step Forward in Multimodal LLM Alignment中国科学院自动化研究所 · 2025年



