README.md

---
pretty_name: ADOPD2026
license: cc-by-nc-nd-4.0
language:
- en
- zh
- ja
- multilingual
task_categories:
- object-detection
- image-segmentation
- image-classification
- image-to-text
size_categories:
- 100K<n<1M
tags:
- document-ai
- document-understanding
- visual-grounding
- ocr
- instance-segmentation
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: validation
    path: data/validation-*
  - split: test
    path: data/test-*
---

<div align="center">
  <img src="assets/adopd-headline.png" width="100%" alt="ADOPD2026: From Page Decomposition to Grounded Document Reasoning">
  <p>
    <a href="https://sichenzhu.github.io/thinking-with-anchors/">Project</a> |
    <a href="https://github.com/SichenZhu/ADOPD2026">Code</a> |
    <a href="https://github.com/SichenZhu/ADOPD2026/tree/main/release_code/model_zoo">Model Zoo</a>
  </p>
</div>

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 retains independent annotation
layers and their provenance boundaries. Human entity polygons are canonical
segmentation targets; grouped OCR blocks are canonical text-box targets; VLM
annotations add semantic labels to human masks; legacy masks remain available
as a separate historical layer.

## 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 | 76.06 GB |

| Annotation layer | Count | Intended use |
| --- | ---: | --- |
| Human entity masks | 1,222,592 | canonical polygon localization and segmentation |
| Kimi-K2.5 semantic mask annotations | 1,132,311 | region tagging and semantic filtering |
| Grouped OCR blocks | 1,246,351 | text-box detection and grouping |
| Human-tagged OCR blocks | 134,248 | semantic text-role and foreground/background analysis |
| Legacy masks | 949,193 | independent historical Labelbox layer |
| Legacy role masks | 79,668 | historical semantic/visual-role layer |

![ADOPD2026 annotation scale](assets/annotation-scale.svg)

## Annotation Layers

![ADOPD2026 annotation layers](assets/annotation-layers.svg)

The layers are related by page identity, not by positional list index:

- `human_annotated_masks` contains the canonical human polygon geometry. Its
  stable `mask_index` joins an optional semantic `vlm_annotation`.
- `ocr.grouped_blocks` contains human-grouped OCR geometry and text. It is the
  supervision used by released text detection checkpoints.
- `ocr.human_tagged_blocks` is a second text layer with semantic role and
  foreground/background labels.
- `legacy_masks` and `legacy_role_masks` are independent annotations. They must
  not be zipped or positionally matched with `human_annotated_masks`.

## Semantic Region Labels

Human masks with a successful VLM annotation use 13 public values:

`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`.

![ADOPD2026 semantic label distribution](assets/semantic-labels.svg)

The model name stored in the release is the public name `moonshotai/Kimi-K2.5`.
Masks without a validated annotation keep their geometry and omit the semantic
payload; they are not assigned a guessed class.

## Loading The Dataset

### Hugging Face Datasets

```python
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["ocr"]["grouped_blocks"]
```

For quick inspection without downloading all shards:

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

### Direct Parquet Access

```python
import duckdb
from huggingface_hub import HfFileSystem

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

### Released Training Adapters

The companion [ADOPD2026 code repository](https://github.com/SichenZhu/ADOPD2026)
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.

### Identity And Image

| Field | Type | Description |
| --- | --- | --- |
| `sample_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_sha256` | `string` | Non-reversible hash of the source URL for identity checks |
| `language` | `string` | Primary annotation language |

### Captions And Global Tags

| Field | Description |
| --- | --- |
| `captions.human` | Human-authored page caption when available |
| `captions.gpt4v` | GPT-4V page caption when available |
| `captions.blip2_opt_6_7b` | Optional BLIP2 OPT-6.7B caption |
| `captions.preferred_source` | Which non-empty caption source was selected by metadata: `human`, `gpt4v`, or `source_text` |
| `global_tags.human` | Human document-level tags |
| `global_tags.clip` | CLIP tag/probability records |
| `global_tags.ram_attributes` | RAM-generated visual attributes |

### OCR

| Field | Description |
| --- | --- |
| `ocr.source_text` | Source plain text associated with the page |
| `ocr.paddleocr_v4_raw` | Retained PaddleOCR v4 output text |
| `ocr.internvl3_78b_text` | Retained InternVL3-78B OCR transcription |
| `ocr.grouped_blocks` | Canonical grouped text blocks used for Doc2Box |
| `ocr.human_tagged_blocks` | Optional human-tagged text blocks with semantic role and visual layer |

Each `grouped_blocks` item contains `block_index`, `group_index`, `polygon`,
`bbox_xyxy`, `text`, `confidence`, and `source_member_indices`. Each
`human_tagged_blocks` item contains the same stable indices and geometry plus
`text_role` and `layer`.

### Human And Legacy Masks

| Field | Description |
| --- | --- |
| `human_annotated_masks` | Canonical list of human entity polygons |
| `legacy_masks` | Independent historical mask/label list |
| `legacy_role_masks` | Historical masks augmented with coarse semantic/visual role |
| `legacy_mask_summary` | Page-level historical `primary_label` and derived `label_family` |
| `legacy_annotation_project` | Source project name for the legacy layer |

Each human mask has `mask_index`, `polygons`, `source_label`, `bbox_xyxy`,
`area_fraction`, and nullable `vlm_annotation`. A VLM annotation contains only
`label`, `model`, and `short_reason`. Each legacy mask contains `polygons`,
`label`, `bbox_xyxy`, and `area_fraction`; a legacy role mask additionally has
`role`.

### Quality, Document Metadata, And Statistics

| Field | Description |
| --- | --- |
| `quality.unsafe_probability` | Estimated unsafe-content probability |
| `quality.watermark_probability` | Estimated watermark probability |
| `quality.caption_image_similarity` | Caption-image similarity score |
| `document_metadata.language_group` | Coarse language group |
| `document_metadata.language_detail_group` | Fine-grained language group |
| `document_metadata.layout_profile` | Coarse page layout profile |
| `document_metadata.text_visual_profile` | Text-versus-visual content profile |
| `document_metadata.taxonomy_cluster_id` | Stable taxonomy cluster identifier |
| `document_metadata.taxonomy_cluster_name` | Human-readable taxonomy cluster |
| `document_metadata.taxonomy_assignment_tag` | Assignment tag retained from taxonomy analysis |
| `statistics` | Page-level size, object, text, polygon, and complexity features |

`statistics` contains `aspect_ratio`, `megapixels`, `object_count`,
`bbox_count`, `polygon_count`, `vertex_count`, `ocr_block_count`,
`unique_mask_label_count`, `small_bbox_ratio`, `bbox_area_std`,
`text_rich_score`, `visual_rich_score`, `bbox_complexity_score`,
`segmentation_complexity_score`, and aggregate `complexity_score`.

## 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.
- The full staging audit counted all 120,000 rows and decoded representative
  first/last rows from every split.
- The Hugging Face viewer was validated against the published Parquet commit.
- Numeric source label IDs, local paths, source URLs, run IDs, retries, prompts,
  endpoint names, and other annotation-process state are intentionally omitted.

The source contained 1,062 `Class-Delete` raster-only items without renderable
polygon geometry. They 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.
Legacy layers have different collection semantics and are not interchangeable
with canonical human masks.

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

ADOPD2026 is released under
[CC BY-NC-ND 4.0](https://creativecommons.org/licenses/by-nc-nd/4.0/).
It is intended for non-commercial research use. Review the license before
redistributing data or derived artifacts.

## Citation

```bibtex
@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}
}
```

```bibtex
@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}
}
```