BaFCo — Bangla Form Understanding

Real Bangladeshi government forms with manual annotations for two tasks:

• bafco_dla/ — Document Layout Analysis: layout bounding boxes, inter-box relations, and form-level labels; images partitioned by domain.
• bafco_kie/ — Key Information Extraction: key–value pairs; forms split by language (English / Bangla).

Each subset is self-contained: annotations in all_annotations.json, images under the subset folder, and a README.md documenting the schema. Image references (local_images) are relative and resolve after download.

Load

from huggingface_hub import snapshot_download
import os, json

root = snapshot_download("Mausul/bafco", repo_type="dataset")

dla = json.load(open(os.path.join(root, "bafco_dla", "all_annotations.json"), encoding="utf-8"))
img = os.path.join(root, "bafco_dla", *dla[0]["local_images"][0].split("/"))   # -> a real file

release_loader.py (repo root) adapts the annotations into the structures the evaluation pipeline expects:

import sys; sys.path.insert(0, root)
import release_loader
dla_targets = release_loader.load_dla_targets(os.path.join(root, "bafco_dla", "all_annotations.json"))
kie_forms   = release_loader.load_kie_forms(os.path.join(root, "bafco_kie", "all_annotations.json"))

See each subset's README.md for the full schema.

Code

Full benchmark pipeline (inference, evaluation, post-processing): coming soon

License

cc-by-nc-4.0