Dataset · MAmmoTH-VL
12M visual instructions for multimodal models
MAmmoTH-VL/MAmmoTH-VL-Instruct-12M
Image-instruction pairs for training visual assistants in English.
What's inside
The dataset contains 12M visual instructions (images and text) in English. Data was gathered via a scalable pipeline: manual source collection, rewriting using MLLM/LLM, and filtering by the same MLLM judge. The card shows transformation examples for mathematics and natural sciences.
What you can do with it
- SFT MAmmoTH-VL-8B
- LoRA fine-tuning for visual question answering
- Evaluation of multimodal models on math and science
- Tasks
- Visual question answering, Question answering
- Languages
- English
- Size
- 10M–100M rows
- Downloads this month
- 15.4K
- Likes
- 67
Models trained on it
- HuggingFaceTB/SmolVLM2-500M-Video-InstructImage text → text 1.7M
- HuggingFaceTB/SmolVLM2-2.2B-InstructImage text → text 194.6K
- HuggingFaceTB/SmolVLM2-256M-Video-InstructImage text → text 44.7K
- HuggingFaceTB/SmolVLM2-2.2B-BaseImage text → text 2.4K
- mradermacher/SmolVLM2-2.2B-Instruct-i1-GGUFVideo text → text 1.1K
- mradermacher/SmolVLM2-256M-Video-Instruct-i1-GGUF 824
- mradermacher/SmolVLM2-500M-Video-Instruct-i1-GGUF 737
- mradermacher/SmolVLM2-2.2B-Instruct-GGUFVideo text → text 368
How to load it
from datasets import load_dataset
ds = load_dataset("MAmmoTH-VL/MAmmoTH-VL-Instruct-12M")