RUC-AIBOX/Virgo-Visual-Long-Thought-Dataset

Virgo-Visual-Long-Thought是一个用于训练Virgo-7B模型的长思考视觉推理数据集，基于8个视觉问答数据集构建，包括几何数据集、表格和图形数据集以及对象数据集。数据集包含原始问题、答案、模型生成的响应、用于训练的最终数据、图像路径和正确性得分。Virgo-Visual-Long-Thought is a long thought visual reasoning dataset for training the Virgo-7B model, constructed based on 8 visual question answering datasets including geometry datasets, table and figure datasets, and object datasets. The dataset contains the original question, answer, model-generated responses, final training data, image paths, and correctness scores.

搜集汇总

背景与挑战

背景概述

Virgo-Visual-Long-Thought-Dataset是一个用于训练多模态大语言模型（Virgo-7B）的视觉长思维推理数据集。它基于8个视觉问答数据集（如几何、图表、对象等）构建，每个样本包含原始问题、答案以及由Virgo-72B生成的多条思维链响应，并筛选出正确且最短的响应用于模型训练。该数据集旨在复现类似o1的推理能力，提升模型在视觉推理任务中的表现。