Datasets:

open-index
/

open-html

	---
	license: odc-by
	task_categories:
	- text-generation
	- feature-extraction
	- text-classification
	language:
	- en
	- mul
	pretty_name: OpenHTML
	size_categories:
	- 1M<n<10M
	tags:
	- common-crawl
	- web-crawl
	- html
	- text
	- metadata
	configs:
	- config_name: default
	data_files:
	- split: train
	path: data//
	- config_name: CC-MAIN-2026-12
	data_files:
	- split: train
	path: data/CC-MAIN-2026-12/*
	---

	# OpenHTML

	> Raw HTML from the web with rich structured metadata — ready for training, retrieval, and analysis

	## What is it?

	OpenHTML is a large-scale web dataset built from [Common Crawl](https://commoncrawl.org). Common Crawl is a non-profit that crawls the web and freely provides its archives and datasets to the public — see [their latest crawl announcement](https://commoncrawl.org/blog/march-2026-crawl-archive-now-available) for details on the source data. Every page goes through a pipeline that extracts the raw HTML body along with structured metadata from WARC records, HTTP response headers, and HTML `<head>` tags, then packages everything into Parquet files with 24 columns.

	The dataset currently includes crawl CC-MAIN-2026-12 with 197,357 documents across 10 shards. Processed 34.3 GB of raw HTML into 34.3 GB of stored body text — 6.5 GB as Parquet (Zstd). We plan to add more snapshots over time.

	OpenHTML is released under the Open Data Commons Attribution License (ODC-By) v1.0, the same license used by Common Crawl.

	## What is being released?

	Each Common Crawl WARC file (~1 GB of compressed HTML) becomes one Parquet shard. The shards live under a crawl-specific directory so multiple snapshots can coexist:

	```
	data/
	CC-MAIN-2026-12/
	00000.parquet
	00001.parquet
	...
	```

	Every row in a Parquet file is one web page with 24 columns of metadata. Each row includes the `warc_record_id` and `warc_date` fields parsed from the original WARC headers, so you can trace any document back to its source record. We also extract HTTP response headers (`content_type`, `charset`, `content_language`, `http_server`, `http_last_modified`) and HTML `<head>` metadata (`title`, `description`, `og:title`, `og:description`, `og:image`, `og:type`, `canonical_url`, `html_lang`). The URL is decomposed into `host`, `domain` (eTLD+1), `path`, and `query`.

	## How to download and use OpenHTML

	### Using `datasets`

	```python
	from datasets import load_dataset

	# stream the entire dataset
	ds = load_dataset("open-index/open-html", name="CC-MAIN-2026-12", split="train", streaming=True)
	for doc in ds:
	print(doc["url"], doc["title"], len(doc["body"]))

	# load a single shard into memory
	ds = load_dataset(
	"open-index/open-html",
	data_files="data/CC-MAIN-2026-12/00000.parquet",
	split="train",
	)
	```

	### Using `huggingface_hub`

	```python
	from huggingface_hub import snapshot_download

	folder = snapshot_download(
	"open-index/open-html",
	repo_type="dataset",
	local_dir="./open-html/",
	allow_patterns="data/CC-MAIN-2026-12/*",
	)
	```

	For faster downloads, install `pip install huggingface_hub[hf_transfer]` and set `HF_HUB_ENABLE_HF_TRANSFER=1`.

	### Using DuckDB

	```sql
	SELECT url, title, domain, html_lang, html_length
	FROM read_parquet('hf://datasets/open-index/open-html/data/CC-MAIN-2026-12/*.parquet')
	WHERE domain = 'wikipedia.org'
	LIMIT 10;
	```

	```sql
	-- Top domains by page count
	SELECT domain, COUNT(*) as pages, AVG(html_length) as avg_html
	FROM read_parquet('hf://datasets/open-index/open-html/data/CC-MAIN-2026-12/*.parquet')
	GROUP BY domain
	ORDER BY pages DESC
	LIMIT 20;
	```

	```sql
	-- Pages with Open Graph metadata
	SELECT url, og_title, og_description, og_image
	FROM read_parquet('hf://datasets/open-index/open-html/data/CC-MAIN-2026-12/*.parquet')
	WHERE og_title != '' AND og_image != ''
	LIMIT 10;
	```

	# Dataset card for OpenHTML

	## Dataset Description

	- Homepage and Repository: [https://huggingface.co/datasets/open-index/open-html](https://huggingface.co/datasets/open-index/open-html)
	- Point of Contact: please create a discussion on the Community tab
	- License: Open Data Commons Attribution License (ODC-By) v1.0

	## Dataset Structure

	### Data Instance

	The following is an example row from the dataset:

	```json
	{
	"url": "https://example.com/article/interesting-topic",
	"warc_date": "2026-03-05T07:14:58Z",
	"warc_record_id": "a1b2c3d4-e5f6-7890-abcd-ef1234567890",
	"warc_filename": "CC-MAIN-20260305070756-20260305100756-00000.warc.gz",
	"http_status": 200,
	"content_type": "text/html",
	"charset": "utf-8",
	"content_language": "en",
	"http_server": "nginx",
	"http_last_modified": "Tue, 04 Mar 2026 12:00:00 GMT",
	"host": "example.com",
	"domain": "example.com",
	"path": "/article/interesting-topic",
	"query": "",
	"html_lang": "en",
	"title": "Interesting Topic - Example",
	"description": "A fascinating article about interesting topics.",
	"og_title": "Interesting Topic",
	"og_description": "A fascinating article about interesting topics.",
	"og_image": "https://example.com/images/topic.jpg",
	"og_type": "article",
	"canonical_url": "https://example.com/article/interesting-topic",
	"html_length": 48210,
	"body": "<!DOCTYPE html><html lang=\"en\"><head>..."
	}
	```

	### Data Fields

	\| Column \| Type \| Description \|
	\|---\|---\|---\|
	\| `url` \| string \| Full URL of the crawled page \|
	\| `warc_date` \| string \| Crawl timestamp from the WARC record (RFC 3339) \|
	\| `warc_record_id` \| string \| UUID from the WARC-Record-ID header, for source traceability \|
	\| `warc_filename` \| string \| Source WARC file basename from Common Crawl \|
	\| `http_status` \| int32 \| HTTP response status code (always 200 in this dataset) \|
	\| `content_type` \| string \| Content-Type from the HTTP response (always starts with `text/html`) \|
	\| `charset` \| string \| Character encoding from the Content-Type header (e.g., `utf-8`, `iso-8859-1`) \|
	\| `content_language` \| string \| Content-Language HTTP header (e.g., `en`, `de`, `fr`) \|
	\| `http_server` \| string \| Server software from the HTTP response (e.g., `nginx`, `Apache`) \|
	\| `http_last_modified` \| string \| Last-Modified HTTP header — when the page was last changed \|
	\| `host` \| string \| Lowercase hostname extracted from the URL (e.g., `www.example.com`) \|
	\| `domain` \| string \| Registered domain (eTLD+1) — groups subdomains together (e.g., `example.com`) \|
	\| `path` \| string \| URL path component (e.g., `/article/interesting-topic`) \|
	\| `query` \| string \| URL query string, if any (e.g., `page=2&sort=date`) \|
	\| `html_lang` \| string \| Language attribute from `<html lang="...">` tag \|
	\| `title` \| string \| Page title from `<title>` tag in `<head>` \|
	\| `description` \| string \| Meta description from `<meta name="description">` \|
	\| `og_title` \| string \| Open Graph title from `<meta property="og:title">` \|
	\| `og_description` \| string \| Open Graph description from `<meta property="og:description">` \|
	\| `og_image` \| string \| Open Graph image URL from `<meta property="og:image">` \|
	\| `og_type` \| string \| Open Graph type from `<meta property="og:type">` (e.g., `article`, `website`) \|
	\| `canonical_url` \| string \| Canonical URL from `<link rel="canonical">` — the page's preferred URL \|
	\| `html_length` \| int64 \| Byte length of the raw HTML body in bytes \|
	\| `body` \| string \| Raw HTML body (full content, no truncation) \|

	### Data Splits

	The default subset includes all available data across all crawl snapshots. You can also load a specific crawl by using its ID as the config name (e.g. `CC-MAIN-2026-12`).

	## Dataset Creation

	### Curation Rationale

	Most open web datasets either release raw text (losing structure) or processed markdown (losing metadata). OpenHTML takes a different approach: it preserves the raw HTML alongside 24 columns of structured metadata extracted from WARC headers, HTTP response headers, and HTML `<head>` tags. This lets you:

	- Train models on raw web content with full context
	- Filter by language, domain, content type, or Open Graph metadata
	- Analyze web structure, server software distribution, or charset usage
	- Trace every document back to its exact WARC source record

	### Source Data

	The source data consists of web pages crawled by the [Common Crawl](https://commoncrawl.org) foundation. Common Crawl archives billions of pages across the public web and makes the raw WARC files freely available on Amazon S3.

	### Data Processing Steps

	The processing pipeline runs as a single-pass extraction:

	1. Download raw .warc.gz files from Common Crawl S3 (each file is roughly 1 GB compressed)
	2. Filter to keep only HTTP 200 responses with a `text/html` content type, discarding images, scripts, redirects, and error pages
	3. Parse HTTP response headers to extract `content_type`, `charset`, `content_language`, `server`, and `last_modified`
	4. Decompose the URL into `host`, `domain` (eTLD+1 via the Public Suffix List), `path`, and `query`
	5. Extract HTML `<head>` metadata using a streaming tokenizer: `title`, `description`, Open Graph tags (`og:title`, `og:description`, `og:image`, `og:type`), `canonical_url`, and `html_lang`
	6. Store the full HTML body (no truncation — `html_length` matches `body` size)
	7. Export directly to Apache Parquet with Zstd compression, 100,000 rows per row group

	No intermediate files are created — the pipeline streams from compressed WARC through extraction directly into Parquet. Pages that produce empty HTML bodies are dropped.

	### Compression Ratios

	Numbers below are actual measurements summed across all 10 files of CC-MAIN-2026-12 (197,357 pages total), projected to the full crawl of 100,000 WARC files.

	\| Stage \| 10 files (measured) \| 100,000 files (projected) \| Reduction \|
	\|---\|---\|---\|---\|
	\| Raw WARC (.warc.gz, downloaded) \| ~8.1 GB \| ~79.2 TB \| — \|
	\| HTML extracted (uncompressed) \| 34.3 GB \| ~335.2 TB \| — \|
	\| Body stored (full HTML) \| 34.3 GB \| ~335.2 TB \| -0.0% vs HTML \|
	\| Final Parquet (Zstd) \| 6.5 GB \| ~63.8 TB \| -81.0% vs body \|

	The body column stores the full raw HTML. Parquet with Zstd then compresses the data further. End to end: ~8.1 GB of raw gzipped WARCs becomes 6.5 GB of Parquet — a 19.5% total reduction — containing 197,357 web pages with full metadata.

	### Processing Times

	Pipeline timings across 10 shards of CC-MAIN-2026-12:

	```
	Download (raw WARC) ████████████████████████ 1h 29m 47s
	Extract (WARC → HTML + metadata) ███████████████████████░ 1h 28m 15s
	Publish (HuggingFace upload) ███░░░░░░░░░░░░░░░░░░░░░ 12m 58s
	```

	### Dataset Charts

	![Total size: HTML vs Body vs Parquet](charts/totals_chart.png)

	![Pipeline stage durations](charts/timing_chart.png)

	### Personal and Sensitive Information

	No additional PII filtering is applied beyond what Common Crawl provides. As the dataset is sourced from the public web, it is likely that some personally identifiable information is present. If you find your own PII in the dataset and would like it removed, please open an issue on the repository.

	## Considerations for Using the Data

	### Social Impact

	By releasing both the dataset and the full processing pipeline, we aim to lower the barrier to training and evaluating language models on high quality web data. Researchers and practitioners who cannot afford to run their own Common Crawl processing pipelines can use OpenHTML directly.

	### Discussion of Biases

	OpenHTML inherits the biases present in Common Crawl and the public web at large. The filtering step keeps only `text/html` pages, which may underrepresent content served as other content types. We have not applied any machine-learning-based quality or toxicity filters, as such filters have been shown to disproportionately remove content from certain dialects and communities.

	### Known Limitations

	The full HTML body is stored without truncation. Very large pages (e.g., pages with inline data URIs) will increase shard sizes. The `html_length` field reflects the exact body size in bytes.

	Metadata extraction scans only the `<head>` section for performance. Pages that place `<meta>` or `<title>` tags in the `<body>` will have missing metadata.

	## Additional Information

	### Licensing

	The dataset is released under the Open Data Commons Attribution License (ODC-By) v1.0. The use of this dataset is also subject to [Common Crawl's Terms of Use](https://commoncrawl.org/terms-of-use). The original content remains subject to the rights and terms of its respective publishers.

	### Contact

	Please open a discussion on the [Community tab](https://huggingface.co/datasets/open-index/open-html/discussions) for questions, feedback, or issues.