Nemotron-VLM-Dataset V2: 9 Million Vision-Language Reasoning Samples

NVIDIA released this collection of approximately 9 million vision-language samples in late 2025. It focuses on document understanding, visual question answering, and video-to-text tasks across multiple languages.

Use Cases

• Training models for financial document analysis using the fintabnet instruction subset
• Developing visual reasoning capabilities using the nights_cot Chain-of-Thought entries
• Fine-tuning video-to-text models for temporal visual understanding

Strengths

• Scale of approximately 9 million samples
• Includes specialized nights_cot reasoning paths
• Multilingual support for global vision-language tasks

Limitations

• Potential for label noise as evidenced by the need to filter broken <think> entries in recent updates
• High storage and compute requirements due to the 9 million multimodal record count

Provenance

Source

NVIDIA

Freshness

Last updated December 2025

Geography

Global