DocILE

Fits invoice and purchase-order extraction agents. Borrow its 55-class field taxonomy, its line-item grouping task, and its scorer that matches predicted fields to ground truth by bounding-box overlap and reports precision, recall and F1 per field.

What it asks the agent to do

• locate and extract key fields (amounts, dates, identifiers) from an invoice or order
• extract each line item from a table on a business document
• handle a document layout never seen in training

What to borrow from it

grading designcase templatetaxonomy

Caveats

The 2023 competition has ended; the data download needs a token from docile.rossum.ai and the dataset license is not stated in the paper; only 6.7k annotated documents, padded with 100k synthetic ones; built for layout models, not for agent workflows.

Details

Links

• Paper
• Github