BaFCo: A Document Understanding Benchmark for Complex Bangla Form Comprehension (ECCV 2026)

Real Bangladeshi government forms with manual annotations for two tasks: Document Layout Analysis (DLA) and Key Information Extraction (KIE).

This is the dataset and its schema. The full benchmark pipeline (inference, evaluation, and post-processing) lives in the code repository: github.com/mausulazad/BaFCo. Paper: arXiv:2607.05614.

BaFCo curates 200 complex multi-page government forms across 316 pages and 15 domains. It provides 16,382 layout entities and 8,771 inter-field relationships under a fine-grained 26-entity schema (with a separate 5-type coarse set) for DLA, and 1,926 key-value pairs across 156 forms for KIE. The forms come from sectors such as agriculture, education, banking, and land management.

Subsets

• bafco_dla/: Document Layout Analysis. Layout bounding boxes, inter-box relations, and form-level labels. Images are partitioned by domain.
• bafco_kie/: Key Information Extraction. Key-value pairs. Forms are split by language (English and 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

Download the dataset, then read a subset's annotations. Each record is one form; image paths live in local_images as repo-relative strings.

from huggingface_hub import snapshot_download
import os, json

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

forms = json.load(open(os.path.join(root, "bafco_dla", "all_annotations.json"), encoding="utf-8"))

# Resolve the first page of the first form to an OS-native file path.
rel = forms[0]["local_images"][0]                      # e.g. "images/<domain>/form_<id>/<page>.jpg"
page_path = os.path.join(root, "bafco_dla", *rel.split("/"))

The KIE subset loads the same way from bafco_kie/all_annotations.json (or the en_ / bn_ language splits). Each subset's README.md documents the full annotation schema.

Running the benchmark

Inference, evaluation, and post-processing are in the code repository, which reads these files through a release_loader.py adapter. See github.com/mausulazad/BaFCo to reproduce the results.

License

Code: MIT, Dataset: CC-BY-NC-4.0.