JamesGoGo/

DeepVision-103K

A Visually Diverse, Broad-Coverage, and Verifiable Mathematical Dataset for Multimodal Reasoning

Training on DeepVision-103K yields top performance on both multimodal mathematical reasoning and general multimodal benchmarks:

Average Performance on multimodal math and general multimodal benchmarks. Training on DeepVision-103K elicits more efficient reasoning.

Benchmark Qwen3-VL-8B-Instruct (Acc / Tokens) Qwen3-VL-8B-DeepVision (Acc /… See the full description on the dataset page: https://huggingface.co/datasets/JamesGoGo/DeepVision-103K.