Access to ADOPD2026 is conditioned on acceptance of the
ADOPD2026 Dataset Research Use Terms.
The Dataset is available only for permitted non-commercial research and educational use.

ADOPD2026: From Page Decomposition to Grounded Document Reasoning

Project | Paper | Code | Model Zoo

ADOPD2026 is a 120,000-page document understanding dataset with original high-resolution images and complementary supervision for text detection, semantic region grounding, instance segmentation, page description, and document-level analysis. It is the public data release accompanying Thinking with Anchors: Grounded and Efficient Document Reasoning.

Unlike a flattened detection export, each row keeps the canonical supervision needed by the released tasks. Human-annotated polygons are segmentation targets, human-annotated text blocks are text-box targets, and semantic mask annotations add region-level labels.

Dataset Summary

Pages	Splits	Original images	Parquet shards	Release size
120,000	80k train / 20k validation / 20k test	JPEG bytes, no resize or re-encode	74	75.05 GB

Annotation layer	Count	Intended use
Human-annotated masks	1,222,592	canonical polygon localization and segmentation
Semantic mask annotations	1,132,311	region tagging and semantic filtering
Human-annotated text blocks	1,246,351	text-box detection and grouping

Annotation Layers

human_annotated_masks contains polygons annotated by human reviewers. Its stable mask_index joins an optional semantic vlm_annotation.
human_annotated_text_blocks contains OCR-derived regions grouped into text blocks by human reviewers. These geometries supervise the released text-detection checkpoints; legacy OCR text and detector scores are not part of this release.

Both localization layers are human annotations: masks capture document entities, while human-annotated text blocks capture text regions.

Semantic Region Labels

Human masks with a released semantic annotation use 13 labels:

Background Image, Brand Logo, Chart / Graph, Color Block, Decorative / Pattern Graphic, Dialog Box, Icon, Illustration / Artwork, Line / Divider, Other / Not Target, Photograph, Table, and Text Block / Content.

Semantic mask labels are model-assisted annotations that passed structural validation before the final join. Masks without a released semantic annotation keep their geometry and omit the semantic payload; they are not assigned a guessed class.

Loading The Dataset

Hugging Face Datasets

from datasets import load_dataset

dataset = load_dataset("adopd/adopd2026")
sample = dataset["validation"][0]

image = sample["image"]
human_masks = sample["human_annotated_masks"]
text_blocks = sample["human_annotated_text_blocks"]

For quick inspection without downloading all shards:

stream = load_dataset(
    "adopd/adopd2026",
    split="validation",
    streaming=True,
)
sample = next(iter(stream))

Direct Parquet Access

import duckdb
from huggingface_hub import HfFileSystem

duckdb.register_filesystem(HfFileSystem())
duckdb.sql("""
    SELECT id, width, height, language
    FROM 'hf://datasets/adopd/adopd2026/data/validation-*.parquet'
    LIMIT 10
""").show()

Released Training Adapters

The companion ADOPD2026 code repository provides lazy Parquet readers and deterministic COCO, YOLO, and LocateAnything exporters. It also contains training and inference commands for all nine released checkpoints.

Row Schema

Each row represents one document page. Split membership is encoded by the Parquet filename and is not duplicated in the row.

The fields documented below are the complete public top-level schema.

Identity And Image

Field	Type	Description
`id`	`string`	Stable SHA-256-style page identifier used across released tasks
`image`	`Image`	Original JPEG bytes plus a non-absolute display filename
`width`, `height`	`int32`	Original image dimensions in pixels
`image_md5`	`string`	MD5 of the exact embedded JPEG bytes
`source_url`	`string`	Audited source URL; credential-like query strings are removed
`language`	`string`	Primary annotation language

Captions And Global Tags

Field	Description
`caption`	Human-authored page caption when available
`tags`	Human document-level tags

Human-Annotated Text Blocks

Field	Description
`human_annotated_text_blocks`	Human-grouped text-region geometry used for text detection

Each text-block item contains block_index, polygon, and bbox_xyxy. block_index is a page-local, zero-based index in reading annotation order. Legacy OCR text, detector confidence, grouping internals, and source-member indices are not included.

Human Masks

Field	Description
`human_annotated_masks`	Canonical list of human entity polygons

Each human mask has mask_index, polygons, source_label, bbox_xyxy, area_fraction, and nullable vlm_annotation. The vlm_annotation payload contains only label and short_reason. Source labels are normalized to a documented 16-value English vocabulary.

The source-label vocabulary is Refinable / Added, Non-Natural Image, Refinable / Logo, Refinable / Icon, Pre-Annotated Mask, Dialog Box, Refinable / Other, Unknown, Banner, Document, Image, Logo, Icon, Table, Graph, and Unlabeled.

source_label preserves the original mask-annotation workflow category. It is metadata, not the semantic supervision used by the released single-class localization and segmentation checkpoints. Region tagging uses vlm_annotation.label instead.

Safety

Field	Description
`nsfw_probability`	Estimated NSFW or unsafe-content probability retained for safety filtering

Data Integrity

Images are embedded at original resolution and were not re-encoded for this release.
Polygon-derived boxes and area fractions were recomputed and audited.
Split IDs are unique and disjoint across train, validation, and test.
All 120,000 rows and all 74 Parquet shards passed release-schema checks.

Source mask items without renderable polygon geometry are excluded from human_annotated_masks; retained mask_index values can therefore be non-contiguous.

Intended Uses And Limitations

ADOPD2026 supports research in document detection, instance segmentation, visual grounding, semantic region classification, OCR grouping, and grounded reasoning. The semantic mask labels are model-assisted annotations and should not be treated as human ground truth without task-appropriate validation.

Document pages may contain personal, copyrighted, or sensitive material from their original sources. Users are responsible for complying with the dataset license and applicable law, evaluating demographic and language coverage, and avoiding deployment decisions based solely on automated annotations.

License

Access to and use of ADOPD2026 are governed by the standalone ADOPD2026 Dataset Research Use Terms. The Dataset is available only for the permitted non-commercial research and educational uses described in those Terms. It is not distributed under a Creative Commons license.

The custom terms apply to rights controlled by the Dataset Provider in the annotations, selection, arrangement, compilation, and documentation. Rights in third-party source images remain with their respective rights holders. A Dataset access grant does not transfer ownership of or independently license those third-party materials.

Citation

@misc{zhu2026thinkingwithanchors,
  title={Thinking with Anchors: Grounded and Efficient Document Reasoning},
  author={Sichen Zhu and Yuchen Zhu and Wenzhuo Xu and Jason Kuen and Wanrong Zhu and Jing Shi and Xuan Shen and Quanyi Wang and Yiwei Wang and Yujun Cai and Bing Shuai and Qin Zhang and Yongxin Chen and Shilong Liu and Molei Tao and Jiuxiang Gu},
  year={2026},
  eprint={2608.04424},
  archivePrefix={arXiv},
  url={https://arxiv.org/abs/2608.04424}
}

@inproceedings{gu2024adopd,
  title={{ADOPD}: A Large-Scale Document Page Decomposition Dataset},
  author={Jiuxiang Gu and Xiangxi Shi and Jason Kuen and Lu Qi and Ruiyi Zhang and Anqi Liu and Ani Nenkova and Tong Sun},
  booktitle={The Twelfth International Conference on Learning Representations},
  year={2024},
  url={https://openreview.net/forum?id=x1ptaXpOYa}
}

Downloads last month: 116

Total file size:

75 GB

Models trained or fine-tuned on adopd/adopd2026

Collection including adopd/adopd2026

Thinking with Anchors

Collection

11 items • Updated 10 days ago

Paper for adopd/adopd2026

Thinking with Anchors: Grounded and Efficient Document Reasoning

Paper • 2608.04424 • Published 9 days ago