FineVision: 24.3M Vision-Language Samples with 88.9M Dialogue Turns

HuggingFaceM4 released FineVision in October 2025, a collection of 24.3 million samples featuring 17.3 million images and 88.9 million conversational turns. The dataset is designed for training open Vision-Language Models and contains 9.5 billion answer tokens.

Use Cases

• Training multi-turn Vision-Language Models using the 88.9M conversational turns
• Fine-tuning multi-modal instruction followers using the 9.5B answer tokens
• Large-scale image feature extraction using the 17.3M image records

Strengths

• Contains 17.3M images for visual representation learning
• Includes 88.9M conversational turns for complex dialogue training
• Provides 9.5B answer tokens for high-resolution language modeling
• Native support for Polars and Dask for distributed processing

Limitations

• Massive data volume requires high-performance computing infrastructure
• Potential for label noise or inconsistencies across 24.3M samples

Provenance

Source

HuggingFaceM4

Freshness

Updated October 2025