DeepVision-103K: 103,000 Verifiable Multimodal Math Reasoning Pairs

DeepVision-103K contains 103,000 multimodal records focused on verifiable mathematical reasoning, released by skylenage in February 2026. It utilizes image-text pairs to improve the efficiency of vision-language models in solving complex logic problems.

Use Cases

• Fine-tuning vision-language models for multimodal math reasoning using the image and text pairs
• Implementing reinforcement learning for reasoning tasks as indicated by the dataset tags
• Benchmarking model accuracy on verifiable visual-mathematical logic

Strengths

• 103,000 multimodal records
• Verifiable mathematical solutions
• MIT license for open research

Limitations

• Narrow domain focus on mathematics may limit generalizability
• Metadata lacks explicit column-level definitions

Provenance

Source

skylenage (Hugging Face)

Freshness

Last updated February 2026.