Skip to content

Dataset · 5551z

Multimodal dataset with image reasoning

5551z/VisCoR-55K

10k to 100k image-question-answer pairs with contrastive examples and step-by-step rationales.

What's inside

The dataset gathers 10k to 100k multimodal examples (images and text) across five categories: general questions, logic, math, charts/diagrams, and OCR. It contains three components: original image question-answer pairs, contrastive pairs for checking reasoning honesty, and generated answer rationales (using the VC-STaR framework).

What you can do with it

  • Training and SFT of multimodal models (VLM) for visual reasoning
  • Fine-tuning models for working with charts, math, and OCR in images
  • Improving VLM robustness to hallucinations using contrastive pairs
Size
10K–100K rows
Downloads this month
365
Likes
40

Models trained on it

How to load it

from datasets import load_dataset

ds = load_dataset("5551z/VisCoR-55K")