MINT-1T: Open-Source Multimodal Dataset with 1 Trillion Tokens

MINT-1T is an open-source multimodal interleaved dataset containing 1 trillion text tokens and 3.4 billion images, a 10x scale-up from prior open-source collections. It includes previously untapped sources such as PDFs and ArXiv papers and is designed for multimodal pretraining research. The dataset was created by a team from the University of Washington and was last updated on the platform in September 2024.

Use Cases

• Multimodal pretraining based on interleaved image-text sequences.
• Benchmarking model scaling on datasets containing 1 trillion tokens.
• Exploring novel data sources for training based on the inclusion of PDFs and ArXiv papers.
• Training models for image-to-text generation tasks.

Strengths

• Contains 1 trillion text tokens and 3.4 billion images.
• Scales open-source multimodal data by 10x compared to existing datasets.
• Incorporates previously untapped sources like PDFs and ArXiv papers.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count and file formats are unknown, which may limit suitability assessment.
• Data may reflect source bias inherent to the aggregated web and academic sources.

Provenance

Source

University of Washington team (mlfoundations).

Collection Method

Aggregated from multiple open sources including web data, PDFs, and ArXiv papers.

Time Range

Likely includes content up to 2023 based on the title.

Freshness

Last updated 2024-09-19 21:07:39.

Geography

Region tag suggests a US focus, but specific coverage is unknown.