Deepeyes: 47,000 Multimodal Records for Visual Reinforcement Learning

Deepeyes Datasets 47K contains 47,000 multimodal records developed by ChenShawn in 2025 to support the 'DeepEyes' research paper. The data facilitates image-to-text tasks specifically designed to incentivize visual reasoning through reinforcement learning frameworks.

Use Cases

• Training multimodal models to generate text-based reasoning chains from visual inputs
• Implementing reinforcement learning loops to optimize image-to-text outputs based on logical consistency
• Benchmarking visual agent performance on tasks requiring multi-step visual interpretation

Strengths

• 47,000 record count
• Apache 2.0 license allows for broad use
• Directly linked to Arxiv 2505.14362 research for theoretical context

Limitations

• Lack of explicit column documentation in the metadata
• Potential geographic bias toward US-centric data as indicated by repository tags

Provenance

Source

ChenShawn, DeepEyes: Incentivizing 'Thinking with Images' via Reinforcement Learning

Freshness

Last updated May 2025.

Geography

United States