Skip to content

Dataset · m-a-p

PIN-14M Multimodal Document Dataset

m-a-p/PIN-14M

Large corpus of paired and alternating English and Chinese multimodal documents with quality labels.

What's inside

This is a mini-version of the PIN dataset containing 14 million examples in PIN format, totaling about 18.79 TB. Metadata specifies text modality (10K to 100K samples) and two languages: English and Chinese. In the final version of the dataset, the authors updated quality signals for all subsets.

What you can do with it

  • Training multimodal models to understand mixed-content documents
  • Filtering training sets using data quality signals
  • Fine-tuning bilingual (en/zh) systems for complex documents
Languages
English, Chinese
Size
10K–100K rows
Downloads this month
28K
Likes
38

How to load it

from datasets import load_dataset

ds = load_dataset("m-a-p/PIN-14M")