uclanlp/VISCO
收藏Hugging Face2024-12-04 更新2024-12-14 收录
下载链接：
https://hf-mirror.com/datasets/uclanlp/VISCO
下载链接
链接失效反馈官方服务：
资源简介：
VISCO是一个用于评估大型视觉语言模型（LVLMs）批判和纠正能力的基准数据集。该数据集包含1645对问题和LVLM生成的答案，每个答案包含多步推理的思维链，以及5604个步骤的批判注释，显示每个步骤是否正确，并在步骤错误时提供自然语言解释。数据集还包括示例数据的JSON表示，展示了问题、LVLM生成的答案、每个步骤的批判注释等信息。
The VISCO dataset is designed to evaluate the critique and correction capabilities of Large Vision-Language Models (LVLMs). It contains 1645 pairs of questions and LVLM-generated answers, each with a chain-of-thought including multiple reasoning steps. Additionally, the dataset includes 5604 step-wise critique annotations, indicating whether each step is correct or incorrect, with natural language explanations provided for incorrect steps. Each data point in the dataset consists of a question, an LVLM-generated answer, and critique annotations for each step. The annotations include binary critique for the final answer, binary critique for each step, and natural language critique for each step.
提供机构：
uclanlp
