Dataset · TIGER-Lab
Verifiable visual tasks for VLM RL training
TIGER-Lab/ViRL39K
A collection of 38,870 cleaned and verified image-question-answer pairs for training reasoning multimodal models.
What's inside
The ViRL39K dataset contains 38,870 verifiable English tasks in image and text modalities, gathered from new sources and reworked existing sets (Llava-OneVision, MM-Math, DeepScaleR, etc.). All examples underwent cleaning, reformulation, and verification; the VL-Rethinker reasoning model was trained on these exact data.
What you can do with it
- Training multimodal models (VLM) for reasoning using RL
- SFT fine-tuning of models for solving visual math and logic tasks
- Evaluating visual reasoning quality
- Tasks
- Question answering, Image text → text
- Languages
- English
- Downloads this month
- 2K
- Likes
- 43
Models trained on it
- mradermacher/EAN-8B-i1-GGUFReinforcement learning 4.5K
- mradermacher/EAN-7B-i1-GGUFReinforcement learning 3.8K
- mradermacher/EAN-3B-i1-GGUFReinforcement learning 2.5K
- mradermacher/RACRO-7B-CRO-i1-GGUF 1.3K
- mradermacher/EAN-3B-GGUFReinforcement learning 979
- mradermacher/EAN-8B-GGUFReinforcement learning 954
- mradermacher/EAN-7B-GGUFReinforcement learning 925
- mradermacher/RACRO-7B-CRO-GRPO-i1-GGUF 744
How to load it
from datasets import load_dataset
ds = load_dataset("TIGER-Lab/ViRL39K")