Datasets:
source large_stringclasses 6
values | doc_id large_stringlengths 8 72 | doc_images large listlengths 1 379 | doc_ocr large listlengths 1 379 | Q&A large_stringlengths 2 77.9k |
|---|---|---|---|---|
deepform | 429353-collect-files-35870-political-file-2012-federal | [{"bytes":"/9j/4AAQSkZJRgABAQAAAQABAAD/2wBDAAMCAgMCAgMDAwMEAwMEBQgFBQQEBQoHBwYIDAoMDAsKCwsNDhIQDQ4RD(...TRUNCATED) | ["{\"Blocks\": [{\"BlockType\": \"PAGE\", \"Geometry\": {\"BoundingBox\": {\"Height\": 1.0, \"Left\"(...TRUNCATED) | "{\"deepform/19318\": {\"question\": \"What is Gross Amount?\", \"answers\": [{\"value\": \"$38,500.(...TRUNCATED) |
deepform | KSTP-TV_14007068422229 | [{"bytes":"/9j/4AAQSkZJRgABAQAAAQABAAD/2wBDAAMCAgMCAgMDAwMEAwMEBQgFBQQEBQoHBwYIDAoMDAsKCwsNDhIQDQ4RD(...TRUNCATED) | ["{\"Blocks\": [{\"BlockType\": \"PAGE\", \"Geometry\": {\"BoundingBox\": {\"Height\": 1.0, \"Left\"(...TRUNCATED) | "{\"deepform/33681\": {\"question\": \"What is Contract Number?\", \"answers\": [{\"value\": \"21697(...TRUNCATED) |
deepform | 476428-romney-wsoc-337109-10-16-12-13504191312854-_-pdf | [{"bytes":"/9j/4AAQSkZJRgABAQAAAQABAAD/2wBDAAMCAgMCAgMDAwMEAwMEBQgFBQQEBQoHBwYIDAoMDAsKCwsNDhIQDQ4RD(...TRUNCATED) | ["{\"Blocks\": [{\"BlockType\": \"PAGE\", \"Geometry\": {\"BoundingBox\": {\"Height\": 1.0, \"Left\"(...TRUNCATED) | "{\"deepform/26314\": {\"question\": \"What is Gross Amount?\", \"answers\": [{\"value\": \"$19,350.(...TRUNCATED) |
deepform | 489530-mn-for-marriage711128-13511997089969-_-pdf | [{"bytes":"/9j/4AAQSkZJRgABAQAAAQABAAD/2wBDAAMCAgMCAgMDAwMEAwMEBQgFBQQEBQoHBwYIDAoMDAsKCwsNDhIQDQ4RD(...TRUNCATED) | ["{\"Blocks\": [{\"BlockType\": \"PAGE\", \"Geometry\": {\"BoundingBox\": {\"Height\": 1.0, \"Left\"(...TRUNCATED) | "{\"deepform/8390\": {\"question\": \"What is Gross Amount?\", \"answers\": [{\"value\": \"$241,625.(...TRUNCATED) |
vrdu-registration-form | 19950530_JETRO, New York_Dissemination Report_Dissemination Report | [{"bytes":"/9j/4AAQSkZJRgABAQAAAQABAAD/2wBDAAMCAgMCAgMDAwMEAwMEBQgFBQQEBQoHBwYIDAoMDAsKCwsNDhIQDQ4RD(...TRUNCATED) | ["{\"Blocks\": [{\"BlockType\": \"PAGE\", \"Geometry\": {\"BoundingBox\": {\"Height\": 1.0, \"Left\"(...TRUNCATED) | "{\"vrdu_registration_form/1533\": {\"question\": \"What is the Registration Number?\", \"answers\":(...TRUNCATED) |
deepform | 462343-republican-party-of-florida-10-3-12 | [{"bytes":"/9j/4AAQSkZJRgABAQAAAQABAAD/2wBDAAMCAgMCAgMDAwMEAwMEBQgFBQQEBQoHBwYIDAoMDAsKCwsNDhIQDQ4RD(...TRUNCATED) | ["{\"Blocks\": [{\"BlockType\": \"PAGE\", \"Geometry\": {\"BoundingBox\": {\"Height\": 1.0, \"Left\"(...TRUNCATED) | "{\"deepform/8320\": {\"question\": \"What is Gross Amount?\", \"answers\": [{\"value\": \"$104,650.(...TRUNCATED) |
deepform | WBTV_14079658755731 | [{"bytes":"/9j/4AAQSkZJRgABAQAAAQABAAD/2wBDAAMCAgMCAgMDAwMEAwMEBQgFBQQEBQoHBwYIDAoMDAsKCwsNDhIQDQ4RD(...TRUNCATED) | ["{\"Blocks\": [{\"BlockType\": \"PAGE\", \"Geometry\": {\"BoundingBox\": {\"Height\": 1.0, \"Left\"(...TRUNCATED) | "{\"deepform/30727\": {\"question\": \"What is Contract Number?\", \"answers\": [{\"value\": \"11755(...TRUNCATED) |
MP-DocVQA | fgyg0023 | [{"bytes":"/9j/4AAQSkZJRgABAQAAAQABAAD/2wBDAAgGBgcGBQgHBwcJCQgKDBQNDAsLDBkSEw8UHRofHh0aHBwgJC4nICIsI(...TRUNCATED) | ["{\"PAGE\": [{\"BlockType\": \"PAGE\", \"Geometry\": {\"BoundingBox\": {\"Width\": 0.99978607892990(...TRUNCATED) | "{\"MP-VQA/4452\": {\"question\": \"when is the cocktail party?\", \"answers\": [{\"value\": \"6:00 (...TRUNCATED) |
deepform | 416517-collect-files-11289-political-file-2012-non | [{"bytes":"/9j/4AAQSkZJRgABAQAAAQABAAD/2wBDAAMCAgMCAgMDAwMEAwMEBQgFBQQEBQoHBwYIDAoMDAsKCwsNDhIQDQ4RD(...TRUNCATED) | ["{\"Blocks\": [{\"BlockType\": \"PAGE\", \"Geometry\": {\"BoundingBox\": {\"Height\": 1.0, \"Left\"(...TRUNCATED) | "{\"deepform/21352\": {\"question\": \"What is Gross Amount?\", \"answers\": [{\"value\": \"$4,145.0(...TRUNCATED) |
deepform | WAFB_14144513312701 | [{"bytes":"/9j/4AAQSkZJRgABAQAAAQABAAD/2wBDAAMCAgMCAgMDAwMEAwMEBQgFBQQEBQoHBwYIDAoMDAsKCwsNDhIQDQ4RD(...TRUNCATED) | ["{\"Blocks\": [{\"BlockType\": \"PAGE\", \"Geometry\": {\"BoundingBox\": {\"Height\": 1.0, \"Left\"(...TRUNCATED) | "{\"deepform/28651\": {\"question\": \"What is Contract Number?\", \"answers\": [{\"value\": \"11415(...TRUNCATED) |
BoundingDocs-Hard-10K
Vishva007/BoundingDocs-Hard-10K is a curated, visually challenging subset of letxbe/BoundingDocs filtered specifically for real-world noisy scans, archival documents, dense bureaucratic forms, and handwritten entries, while excluding clean synthetic templates (such as FATURA).
It preserves the exact original Parquet schema (source, doc_id, doc_images, doc_ocr, and Q&A) across dedicated train, validation, and test splits (10,000 documents total)—making it ideal for fine-tuning and benchmarking Vision-Language Models (VLMs) on spatially grounded document QA and word-level key information extraction (KIE).
Dataset Composition & Split Distribution
Documents were sampled independently from the upstream train, validation, and test Parquet shards with zero cross-split document leakage:
Subset (source) |
Visual Characteristics | Train Docs | Validation Docs | Test Docs | Total Docs |
|---|---|---|---|---|---|
deepform |
Low-DPI scanned/faxed political ad contracts, invoices & stamps | 3,401 | 400 | 600 | 4,401 |
mp-docvqa |
Multi-page archival industry scans, memos, tables & handwriting | 2,700 | 300 | 500 | 3,500 |
vrdu (vrdu-ad-buy-form, vrdu-registration-form) |
Dense registration filings, nested tables & checkboxes | 1,324 | 165 | 167 | 1,656 |
sp-docvqa |
Faded carbon copies, typewritten memos & handwritten notes | 200 | 26 | 28 | 254 |
funsd |
Noisy faxed/photocopied forms with handwritten entries (100% of upstream splits) | 149 | 19 | 21 | 189 |
| Total Documents | — | 7,774 | 910 | 1,316 | 10,000 |
(Note: vrdu combines vrdu-ad-buy-form [512 train / 64 val / 65 test] and vrdu-registration-form [812 train / 101 val / 102 test]. Each document row contains multiple grounded QA pairs, expanding to ~35,000–45,000 grounded training QA pairs.)
Schema & Coordinate Conventions
Each row represents a multi-page or single-page document with the following fields:
source(string): Original benchmark subset identifier (FUNSD,SP-DocVQA,MP-DocVQA,deepform,vrdu-ad-buy-form,vrdu-registration-form).doc_id(string): Unique document identifier fromletxbe/BoundingDocs.doc_images(list[Image]): List of page images for the document.doc_ocr(list[string]): List of JSON-serialized Amazon Textract OCR outputs (one per page), containing word-levelBlockswith normalized[0.0, 1.0][Left, Top, Width, Height]bounding boxes.Q&A(string): JSON-serialized dictionary mapping question IDs to question text, rephrased question text, extracted answer strings, 1-indexedpagenumbers, and word/answer bounding boxes in[width, height, x, y]format on a[0, 1000]top-left origin scale.
Converting Bounding Boxes to [x1, y1, x2, y2] (0–1000 Scale)
For VLM architectures expecting [x1, y1, x2, y2] normalized coordinates (such as Qwen-VL / Qwen3.5):
def boundingdocs_to_xyxy(box):
"""Converts [width, height, x, y] (0-1000, top-left) to [x1, y1, x2, y2] (0-1000)."""
w, h, x, y = box
return [
max(0, min(1000, int(round(x)))),
max(0, min(1000, int(round(y)))),
max(0, min(1000, int(round(x + w)))),
max(0, min(1000, int(round(y + h)))),
]
Quickstart Usage
import json
from datasets import load_dataset
ds = load_dataset("Vishva007/BoundingDocs-Hard-10K")
sample = ds["train"][0]
print("Source:", sample["source"])
print("Pages :", len(sample["doc_images"]))
qa_dict = json.loads(sample["Q&A"])
first_q = next(iter(qa_dict.values()))
print("Question:", first_q.get("rephrased_question") or first_q["question"])
print("Answers :", first_q["answers"])
Citation & Attribution
This dataset is a filtered derivative of BoundingDocs. Please cite the original authors when using this dataset:
@article{giovannini2025boundingdocs,
title={BoundingDocs: a Unified Dataset for Document Question Answering with Spatial Annotations},
author={Giovannini, Simone and Coppi, Fabio and Gemelli, Andrea and Marinai, Simone},
journal={International Journal on Document Analysis and Recognition (IJDAR)},
year={2025},
doi={10.1007/s10032-025-00563-5},
eprint={2501.03403},
archivePrefix={arXiv},
primaryClass={cs.CL}
}
(Note: While letxbe/BoundingDocs is released under CC-BY-4.0, upstream subsets such as FUNSD and DocVQA carry non-commercial research restrictions, so cc-by-nc-4.0 is specified here to respect original source terms.)
### Summary of Updates Made
* **`size_categories`:** Changed from `1K<n<10K` to `10K<n<100K` (Hugging Face's official bucket for $\ge 10,000$ rows).
* **Total Count & Split Table:** Updated total rows from **9,899** to **10,000**, adjusted `train` counts (`deepform`: 3,401; `mp-docvqa`: 2,700), and corrected the exact `validation` (`vrdu`: 165, `sp-docvqa`: 26, `funsd`: 19) and `test` (`sp-docvqa`: 28) counts to match your verified dataset summary.
* **VRDU Sub-split Footnote:** Added the exact `vrdu-ad-buy-form` and `vrdu-registration-form` breakdown beneath the table.
- Downloads last month
- 100