Dataset · multimodal-reasoning-lab
Zebra-CoT visual reasoning chains
multimodal-reasoning-lab/Zebra-CoT
182k alternating text and image examples for step-by-step visual reasoning.
What's inside
The dataset contains 182,384 samples with logically coherent reasoning chains where text and graphic steps alternate. Examples cover scientific reasoning, 2D and 3D visual reasoning, visual logic, and strategy games. Each example starts with a task condition; other fields in the card are not disclosed.
What you can do with it
- Training VLM for step-by-step multimodal reasoning
- SFT/LoRA for visual logic tasks
- Testing models on 2D/3D reasoning and games
- Tasks
- Multimodal, Image text → text, Visual question answering
- Size
- 100K–1M rows
- Downloads this month
- 29.9K
- Likes
- 84
Models trained on it
- mradermacher/Omni-R1-i1-GGUF 1.8K
- mradermacher/Anole-Zebra-CoT-i1-GGUF 737
- mradermacher/Anole-Zebra-CoT-GGUF 534
- mradermacher/Omni-R1-GGUF 357
- multimodal-reasoning-lab/Bagel-Zebra-CoTMultimodal 69
- array/Qwen2.5-VL-MullGRPOImage text → text 62
- multimodal-reasoning-lab/Anole-Zebra-CoTMultimodal 50
- ModalityDance/Omni-R1Multimodal 42
How to load it
from datasets import load_dataset
ds = load_dataset("multimodal-reasoning-lab/Zebra-CoT")