docile/README.md

DocILE Dataset

DocILE requires manual registration to download.

How to get the data

1. Go to https://docile.rossum.ai/
2. Fill out the Dataset Access Request form: https://forms.gle/poJqGXrxoftWrUsc8
3. You must agree to legal terms for research use
4. Once approved, download the dataset and extract to this directory

Dataset info

• 6,680 annotated documents + 100,000 synthetic documents
• Nearly 1M unlabeled documents for pre-training
• Task: Key Information Extraction (KIE) + Line Item Recognition (LIR)
• Annotations: field types with bounding boxes on document pages
• License: Research use only (requires agreement)

Using with docile-benchmark

pip3 install docile-benchmark
from docile.dataset import Dataset
dataset = Dataset("train", "/path/to/docile")