liyah1616/Nemotron-Image-Training-v3

Nemotron Image Training v3 是一个面向图像的多模态训练数据集合，专为视觉-语言模型设计。作为 Nemotron-VLM-Dataset v2 的扩展，它包含 76 个子数据集，总计约 690 万个样本和 395.6 亿个令牌，覆盖了广泛的图像中心视觉-语言任务，使用人类标注和合成生成的数据混合。每个子数据集提供标准化的对话 JSONL 格式文件以及数据集卡片，描述来源、许可和媒体布局。媒体通过引用（JSONL 中的路径或 URL）提供，需从上游来源下载并本地对齐。该数据集格式兼容 Megatron Energon 风格的多模态加载，适用于商业和非商业用途，由 NVIDIA Corporation 创建于 2026-04-28。

Nemotron Image Training v3 is a collection of image-centric multimodal training data for vision–language models. Similar to Nemotron-VLM-Dataset v2, it was curated as a large-scale, multi-subdataset release where each subset ships a standardized conversation JSONL alongside a dataset card describing sources, licensing, and media layout. Nemotron Image Training v3 expands on v2 with 76 subdatasets totaling approximately 6.9M samples and 39.56B tokens, covering a broad range of image-centric vision–language tasks using a mix of human-annotated and synthetically generated data. Media is provided by reference (paths or URLs inside the JSONL). The dataset is ready for commercial/non-commercial use, owned by NVIDIA Corporation, and created on 2026-04-28.