Skip to content

Dataset · vidore

ColPali training set for document retrieval

vidore/colpali_train_set

127K query-image pairs for training visual document retrieval.

What's inside

This is the ColPali training set: 127,460 query-image pairs. 63% are taken from open academic datasets, 37% are synthetic PDF pages from web crawling with pseudo-questions from Claude-3 Sonnet. The set is entirely English; modalities are images and text.

What you can do with it

  • Train ColPali or LoRA for visual document retrieval
  • Fine-tune a document question answering model on page images
  • Test zero-shot transfer to non-English documents
Tasks
Document question answering, Visual document retrieval
Size
100K–1M rows
Downloads this month
5.3K
Likes
93

Models trained on it

How to load it

from datasets import load_dataset

ds = load_dataset("vidore/colpali_train_set")