Skip to content

Dataset · OpenDataArena

MMFineReason: complex VQA examples for SFT

OpenDataArena/MMFineReason-SFT-123K-Qwen3-VL-235B-Thinking

123k most difficult image-text examples from MMFineReason where Qwen3-VL-4B-Thinking consistently fails.

What's inside

This is a filtered subset of MMFineReason-1.8M: only the most difficult 7% of examples where Qwen3-VL-4B-Thinking fails all 4 attempts. Inside are 123k image and text examples in English for visual question answering and text generation. The format is SFT data.

What you can do with it

  • SFT for vision-language models on complex reasoning examples
  • Fine-tuning VQA models via LoRA on difficult cases
  • Testing VLM robustness on tasks where a weak model fails
Tasks
Visual question answering, Question answering, Text generation
Languages
English
Size
100K–1M rows
Downloads this month
920
Likes
87

How to load it

from datasets import load_dataset

ds = load_dataset("OpenDataArena/MMFineReason-SFT-123K-Qwen3-VL-235B-Thinking")