Skip to content

Dataset · InnovatorLab

Large-scale 46M-example SFT dataset for VLMs

InnovatorLab/Innovator-VL-Instruct-46M

A huge corpus of multimodal instructions (image + text) for full-parameter SFT of vision-language models.

What's inside

The dataset contains from 10 to 100 million image-text pairs in English and continues to be updated by the authors. It is intended for the visual instruction fine-tuning (SFT) stage, which serves as a bridge between multimodal model pre-training and its subsequent alignment via RL.

What you can do with it

  • Full-parameter SFT of vision-language models for a wide range of tasks
  • Improve instruction-following accuracy and reduce hallucinations in multimodal LLMs
  • Prepare a base VLM before the reinforcement learning (RL) stage
Tasks
Image text → text
Languages
English
Size
10M–100M rows
Downloads this month
103.7K
Likes
18

Models trained on it

How to load it

from datasets import load_dataset

ds = load_dataset("InnovatorLab/Innovator-VL-Instruct-46M")