AmbiHTMLBench

Reconstruct document page images as executable HTML.

Code | Browse pages | Download ZIP

AmbiHTMLBench evaluates content, structure, and visual fidelity in image-to-HTML reconstruction. The benchmark covers institutional PDF documents, web-native documentation, newspapers and magazines, and posters or design-heavy pages.

What is included

Material	This release
Reference page images	1,115
Annotated blocks	28,939
Annotation version	GT v5
Evaluation split	test

Each page has an original image and a complete block-annotation JSON file. Images retain their original bytes, resolution and format. The test split is for evaluation; it is not a training split. This package contains reference materials, not model predictions or judge records.

Download and use

Download the complete ZIP and extract it, or fetch the repository:

from huggingface_hub import snapshot_download

snapshot_download(repo_id="Theseus-Lab/AmbiHTMLBench", repo_type="dataset",
                  local_dir="AmbiHTMLBench", ignore_patterns=["archives/*"])

Read a page from the downloaded directory:

import json
from pathlib import Path
from PIL import Image

root = Path("AmbiHTMLBench")
rows = [json.loads(line) for line in (root / "index/pages.jsonl").read_text(encoding="utf-8").splitlines()]
page = rows[0]
image = Image.open(root / page["image_path"])
annotation = json.loads((root / page["annotation_path"]).read_text(encoding="utf-8"))

For the lightweight index only:

from datasets import load_dataset

pages = load_dataset("Theseus-Lab/AmbiHTMLBench", "pages", split="test")

The viewer contains one metadata row per page. Full images and annotations are in the linked page directories; loading the index does not download them. Use a fixed Hub commit revision when reproducing an experiment.

Dataset structure

README.md
release.json
manifest.json
SHA256SUMS
index/pages.jsonl
pages/0001/image.<original-format>
pages/0001/annotation.json
...
annotations/AmbiHTMLBench_GT_v5_1115.json
archives/AmbiHTMLBench.zip

Index rows contain page_id, path, image_path, annotation_path, image_width, image_height, block_count, annotation_version, and source_url. All file paths are relative to the dataset repository root.

Each annotation.json preserves the original page's ID, imageUrl, and contentList. Blocks contain semantic labels in name, text/table/formula content in remark, and pixel coordinates in left, top, width, height where present. Optional fields and Unicode text are retained. The aggregate GT file is an exact source copy for full releases and contains selected pages only for previews. Original imageUrl values remain provenance; use image_path for local images.

SHA256SUMS covers expanded files and the ZIP, excluding itself. manifest.json records source and image identities and expanded-file digests, excluding itself and the ZIP. The ZIP contains the expanded release and manifest, without the outer ZIP or SHA256SUMS.

Evaluation

Install and follow the evaluation code. Use the aggregate GT JSON and separately generated HTML predictions to measure table, text and layout fidelity. For offline reference visualizations, resolve each page's local image through the index. This GT v5 release is separate from the frozen OLS5 subjective experiment's input snapshot.

Materials and intended use

Reference pages include diverse layouts and multilingual text. Source URLs are retained per page; the downloaded images remain usable if those URLs expire. The public annotations are evaluation references and should not be supplied as inputs to models being evaluated. Exposure or training contamination can affect comparisons.

Licensing and citation

Dataset content and presentation figures are CC BY 4.0. Executable code and environment/configuration files are Apache-2.0.