MultiTrust
收藏资源简介:
MultiTrust是由清华大学和北京航空航天大学的研究团队共同创建的综合性数据集,专注于评估多模态大型语言模型(MLLMs)在可信度方面的表现。该数据集包含32个任务,覆盖了真实世界中的多种场景,包括生成和判别任务,以及多模态和纯文本任务。数据集的构建结合了现有的文本、图像和多模态数据集,通过手动和自动方法调整提示、图像和注释,以及通过互联网收集或使用Stable Diffusion等算法合成的新数据集。MultiTrust旨在解决MLLMs在真实应用中的可信度问题,如防止不期望的输出、确保社会影响等,为未来在可信度研究方面的发展提供了标准化和可扩展的评估工具。
MultiTrust is a comprehensive dataset jointly developed by research teams from Tsinghua University and Beihang University (Beijing University of Aeronautics and Astronautics), focusing on evaluating the trustworthiness performance of multimodal large language models (MLLMs). This dataset includes 32 tasks spanning a wide range of real-world scenarios, covering both generative and discriminative tasks, as well as multimodal and pure-text tasks. For its construction, existing text, image, and multimodal datasets are integrated, with prompts, images and annotations adjusted via both manual and automated methods; additionally, new datasets are either collected from the Internet or synthesized using algorithms such as Stable Diffusion. MultiTrust aims to address the trustworthiness issues of MLLMs in real-world applications, such as preventing undesirable outputs and ensuring positive social impacts, providing a standardized and scalable evaluation tool for future trustworthiness-related research.

- 1Benchmarking Trustworthiness of Multimodal Large Language Models: A Comprehensive Study清华大学 · 2024年



