MINT-1T
Multimodal Interleaved Dataset
MINT-1T is one of the largest open-source multimodal interleaved datasets, containing 1 trillion tokens of text and image data. Created by Salesforce Research, it covers various sources such as HTML, PDF, and ArXiv, supporting the training of large multimodal models that can process both text and images simultaneously.
Dataset Highlights
Trillions of multimodal intertwined data, providing a training foundation for the next generation of multimodal large models
Trillion-scale
Contains over 1 trillion tokens of text and image data, making it one of the largest publicly available multimodal intertwined datasets, providing ample data support for large-scale model pre-training.
Diverse Data Sources
Data is meticulously collected from HTML web pages, PDF documents, and ArXiv academic papers, covering various content forms and domain knowledge, ensuring the diversity of training data.
Intertwined Format
Preserves the natural intertwined order of text and images, faithfully restoring the original structure of the documents, allowing the model to learn the contextual relationships between text and images.
Quality Filtering
Employs a multi-stage deduplication and filtering pipeline to systematically remove low-quality, duplicate, and irrelevant content, ensuring high-quality standards for the dataset.
Ready for Training
The data has been preprocessed and formatted, ready for direct use in model training, significantly reducing the initial preparation time for data engineering, allowing researchers to focus on the model itself.
Open License
Uses an open license agreement, supporting academic research and commercial application scenarios, lowering the barriers to data access, and promoting the democratization of multimodal AI research.
Applicable Scenarios
From multimodal pre-training to document understanding, MINT-1T covers various cutting-edge research directions
Multimodal LLM Training
Pre-training large multimodal language models that can understand mixed content of text and images
Document Understanding
Training models on structured documents containing embedded charts and images to enhance document parsing capabilities
Visual Question Answering
Building visual question answering systems that can perform joint reasoning on text and images
Content Generation
Generating natural interleaved text and image content to achieve collaborative creation of text and images
Quick Access to MINT-1T
Quickly obtain the MINT-1T dataset through the Ace Data Cloud API
import requests
# Set your API token
API_TOKEN = "your_api_token_here"
# Request MINT-1T dataset
response = requests.get(
"https://api.acedata.cloud/datasets/mint-1t",
headers={
"Authorization": f"Bearer {API_TOKEN}",
"Accept": "application/json"
},
params={
"source": "html",
"limit": 10
}
)
# Parse the response
data = response.json()
for item in data.get("results", []):
print(f"Text: {item['text'][:100]}...")
print(f"Images: {len(item.get('images', []))} images")
print("---")
3 Steps to Get Started Quickly
From registration to usage, you can start obtaining data in just a few minutes
Register an Account
Visit platform.acedata.cloud to register for an Ace Data Cloud platform account, complete email verification to start using it.
Obtain API Key
Go to the console to create an API Key for accessing the MINT-1T dataset API, supporting calls in various programming languages.
Start Using
Use the API Key to call the dataset interface and obtain multimodal intertwined data in HTML, PDF, or ArXiv sources as needed.
Start Exploring the MINT-1T Dataset
Trillions of multimodal intertwined data, open license, available immediately. Whether you are a multimodal AI researcher or an application developer, MINT-1T can provide strong data support for your projects.