FineReason is a dataset released by OpenDataArena in 2025 for training and evaluating the visual reasoning capabilities of large multimodal models (LMMs). It aims to improve the interpretable and verifiable long-chain reasoning capabilities of models in scenarios such as visual puzzles, games, complex graph reasoning, and STEM (science, technology, engineering, and mathematics) knowledge applications.

This dataset covers various task types, including geometry problems (geometry3k / geo170k), diagram and flowchart comprehension (AI2D), visual reasoning and observation puzzles (visualwebinstruct, etc.). All samples use a uniform data format, including a unique ID, question text, corresponding image, and reasoning-based answer. The dataset is compiled from multiple public subsets and its reasoning chains are distilled using the Qwen3-VL-235B-a22B-thinking model, ensuring that all samples possess a clearly structured, verifiable step-by-step reasoning process and a final solution.

Data composition (continuously expanding):

• BMMR: 42,647 entries
• Euclid30K: 27,111 entries
• ai2d_merged: 2,446 entries
• geo170k (Q&A): 12,101 results
• geometry3k / mathv360k: 9,724 results
• ScienceQA: 6,146 results
• TQA (TextbookQA): 12,565 items
• VisualWebInstruct (filtered): 261,436 results
• MMR1: 1,000 pieces
• VisualSphinx: 3,781 results
• MMOpenR1-8K: 7,428 entries