ImageNet-Think-250K
收藏资源简介:
ImageNet-Think-250K 是一个大规模的合成数据集,旨在帮助开发具有显式推理能力的视觉语言模型(VLMs)。该数据集基于 ImageNet-21k 数据集中的 250,000 张图像构建,为每张图像提供了结构化的思维标记和相应的答案。这些图像由两个最先进的 VLMs(GLM-4.1V-9B-Thinking 和 Kimi-VL-A3B-Thinking-2506)生成。每个图像都伴随着两对思维-答案序列,为训练和评估多模态推理模型提供了一个资源。
ImageNet-Think-250K is a large-scale synthetic dataset designed to facilitate the development of vision-language models (VLMs) with explicit reasoning capabilities. This dataset is constructed from 250,000 images in the ImageNet-21k dataset, and provides structured thought tokens and corresponding answers for each image. These images are generated by two state-of-the-art VLMs: GLM-4.1V-9B-Thinking and Kimi-VL-A3B-Thinking-2506. Each image is paired with two thought-answer sequence pairs, serving as a valuable resource for training and evaluating multimodal reasoning models.

- 1通过阿贡国家实验室 · 2025年



