MINT-1T HTML
This dataset is a component of the MINT-1T project, containing HTML documents with structured text and integrated visual elements. It is used to train models that can process text-image sequences fluidly.
Several billion tokens in interspersed HTML (text + tags), Parquet format
CC-BY-4.0
Description
MINT-1T HTML is a massive excerpt from the MINT-1T project, composed of HTML documents rich in structured text and visual elements. It is intended for training multimodal models capable of reasoning on interspersed sequences (text + image).
What is this dataset for?
- Pre-train large-scale multimodal models (Idefics2, Chameleon, Xgen-MM...)
- Study the abilities of LLMs to manage complex HTML structures in a multimodal context
- Building agents that can interact with rich web content
Can it be enriched or improved?
Yes, you can cross this data with real visual representations, add annotations (types of HTML elements, semantic structure), or even complete them with metadata on integrated styles and scripts. Finer filtering also makes it possible to adapt the training to specific cases (accessibility, e-commerce, etc.).
🔎 In summary
🧠 Recommended for
- Multimodal AI researchers
- Open-source pre-training
- Web browsing projects by LLM
🔧 Compatible tools
- PyTorch
- Hugging Face Datasets
- WebDataSet
- Idefics2
- VLLM multimodal
💡 Tip
Filter documents by HTML content type (forms, tables, paragraphs) to target specific template capabilities.
Frequently Asked Questions
Does this dataset contain images or only textual HTML?
The file contains HTML with an integrated visual structure (referenced images), but the images themselves are not stored directly here.
Is it possible to use it to train a multimodal chatbot model?
Yes, it's a typical dataset use case: it's designed to train agents who can reason about rich documents.
Can it be used as it is or should it be filtered?
It can be used directly, but filtering by type of HTML content can improve relevance according to the training objective.
