d.HTML
收藏资源简介：
# **d.HTML** ## Overview **d.HTML** is a lightweight dataset designed for **Image-to-Text OCR and structured HTML reconstruction tasks**. The dataset pairs document page images with corresponding markup outputs, primarily in **HTML** (and occasionally Markdown-like structures). It is intended for evaluating and training multimodal models that convert visual documents into structured, machine-readable formats. The dataset focuses on preserving **document structure**, including headings, paragraphs, tables, and layout elements. This makes it suitable for benchmarking **vision-language models**, **OCR systems**, and **document understanding pipelines** that require structured output rather than plain text. Repository: https://huggingface.co/datasets/prithivMLmods/d.HTML ## Dataset Characteristics - **Task Type:** Image-to-Text - **Primary Output Format:** HTML - **Language:** English - **Rows:** 110 - **Dataset Size:** ~51.8 MB - **License:** Apache-2.0 Each sample contains: - **image** – a document page image - **response** – the corresponding structured output (HTML / markup text) The responses attempt to preserve structural elements such as: - headings (`<h1>`, `<h2>`, etc.) - paragraphs (`<p>`) - tables (`<table>`, `<tr>`, `<td>`) - semantic layout information - structured document formatting ## Intended Use This dataset can be used for: - Training **vision-language models** for document understanding - Evaluating **OCR systems with structural reconstruction** - Benchmarking **image-to-HTML conversion models** - Testing **layout-aware multimodal models** - Prototyping document parsing pipelines It is particularly useful for experiments involving: - structured OCR - document layout understanding - multimodal LLM evaluation - markup reconstruction from scanned documents ## Example Schema Example dataset structure: ``` { "image": <document_page_image>, "response": "<html><body><p>...</p></body></html>" } ``` The `response` field contains structured markup representing the semantic layout extracted from the document. ## Limitations - The dataset is relatively **small (110 samples)** and should primarily be used for **evaluation, prototyping, or demonstration purposes** rather than large-scale model training. - Some responses may contain **partial structural reconstruction**, depending on document complexity. - The dataset focuses on **markup reconstruction** rather than raw OCR text extraction. ## Citation If you use this dataset in research or experiments, please cite: ``` @misc{prithiv_sakthi_2026, author = { Prithiv Sakthi }, title = { d.HTML (Revision 66330ad) }, year = 2026, url = { [https://huggingface.co/datasets/prithivMLmods/d.HTML](https://huggingface.co/datasets/prithivMLmods/d.HTML) }, doi = { 10.57967/hf/7967 }, publisher = { Hugging Face } } ``` ## Author **Prithiv Sakthi** Hugging Face: https://huggingface.co/prithivMLmods
