---
annotations_creators:
  - expert-annotators
language_creators:
  - found
language:
  - mr
  - hi
  - en
license: cc-by-4.0
task_categories:
  - document-question-answering
  - table-question-answering
  - image-to-text
tags:
  - indic-nlp
  - document-ai
  - ocr-benchmark
  - indian-government-documents
  - document-layout-analysis
  - visual-text-recognition
size_categories:
  - n<1K
---

# IndicGovBench-Doc: Comprehensive Benchmark for Indian Government Documents

**IndicGovBench-Doc** is an evaluation benchmark for document understanding, optical character recognition (OCR), visual layout analysis, and table extraction across Indian government administrative documents (Government Resolutions, Gazettes, Circulars, Certificates).

## Dataset Summary
- **Documents**: 270 full-page official government records.
- **Languages**: Marathi (Mr), Hindi (Hi), English (En), and multi-lingual codeswitching.
- **Annotations**: High-density bounding box segmentation under RFC-style annotation protocol v0.9.
  - Multi-column reading order.
  - Complex nested and irregular table structures.
  - Explicit scope boundaries: `target`, `excluded`, `uncertain`.
  - Word-level spatial anchors aligned with Google Cloud Vision.

## Dataset Structure
Each row in the dataset corresponds to a document page:
- `doc_id`: Unique identifier (e.g. `order_mr_archive_001`).
- `language`: Primary language (`mr`, `hi`, `en`).
- `doc_category`: Document category (`archive`, `current`, `gazette`, etc.).
- `width`, `height`: Canvas pixel dimensions.
- `content_bbox`: Bounding box of all valid content `[min_x, min_y, max_x, max_y]`.
- `ground_truth_blocks`: Structured text blocks with category, polygon, text, and reading order.
- `ground_truth_tables`: Table data with HTML, markdown, rows, cols, and cell grids.
- `metadata`: Outward number, date, department, and scan quality strata.

## Citation
```bibtex
@dataset{indicgovbench2026,
  author = {Orgpedia Foundation},
  title = {IndicGovBench-Doc: Benchmark for Indian Government Document Processing},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/Orgpedia/indicGovBench-Doc}}
}
```