Datasets:

ray0rf1re
/

unclean-web

url stringclasses 7 values	seed_url stringclasses 7 values	scraped_at stringclasses 7 values	og_tags stringclasses 1 value	twitter_tags stringclasses 1 value	schema_org stringclasses 1 value	structured stringclasses 1 value	links stringclasses 1 value	external_links stringclasses 1 value	images stringclasses 1 value	code_blocks stringclasses 1 value	error stringclasses 1 value
https://bloggers.feedspot.com	https://bloggers.feedspot.com	2026-05-20T06:59:08.409451+00:00	{}	{}	[]	{}	[]	[]	[]	[]	fetch_failed_status_0
https://www.tech2geek.net	https://www.tech2geek.net	2026-05-20T06:59:08.410936+00:00	{}	{}	[]	{}	[]	[]	[]	[]	fetch_failed_status_0
https://bloggerspassion.com	https://bloggerspassion.com	2026-05-20T06:59:08.411592+00:00	{}	{}	[]	{}	[]	[]	[]	[]	fetch_failed_status_0
https://www.ontoplist.com	https://www.ontoplist.com	2026-05-20T06:59:08.412147+00:00	{}	{}	[]	{}	[]	[]	[]	[]	fetch_failed_status_0
http://technovisions.wordpress.com	http://technovisions.wordpress.com	2026-05-20T06:59:08.412670+00:00	{}	{}	[]	{}	[]	[]	[]	[]	fetch_failed_status_0
http://www.myheritage.com	http://www.myheritage.com	2026-05-20T06:59:08.413197+00:00	{}	{}	[]	{}	[]	[]	[]	[]	fetch_failed_status_0
http://blog.exoplatform.org	http://blog.exoplatform.org	2026-05-20T06:59:08.413682+00:00	{}	{}	[]	{}	[]	[]	[]	[]	fetch_failed_status_0

🕸️ Unclean Web

Raw, unfiltered web data scraped across a wide variety of sites and packaged for language model pre-training, fine-tuning, and research.

📊 Dataset Statistics

Metric	Value
Total Pages	18,865
Total Token Estimate	24.35M
Unique Sources	27
Schema Version	3.0
Last Updated	2026-05-21 04:40 UTC

🗂️ Available Splits / Subsets

Split	Description	Format
`full` (per batch)	Complete raw scrape, all columns	Parquet / JSONL
`subset_10k_tokens`	~10K token random sample for quick tests	Parquet / JSONL
`subset_1m_tokens`	~1M token curated sample	Parquet / JSONL
`subset_5m_tokens`	~5M token curated sample	Parquet / JSONL
`subset_half`	Random 50% of successful pages	Parquet / JSONL
`subset_openhermes`	OpenHermes 2.5/3 instruction-tuning format	Parquet / JSONL
`subset_dolma`	Dolma / RedPajama / Fineweb pre-training format	Parquet / JSONL
`subset_tokenizer`	Raw 2048-char text shards for tokenizer training	Parquet / JSONL
`merged` (via merge_batches.py)	All batch splits merged into one	Parquet / JSONL

📋 Column Schema

Column	Type	Description
`url`	`string`	Absolute URL of the scraped page
`seed_url`	`string`	Root URL this crawl started from
`depth`	`int64`	BFS depth from seed (0 = seed itself)
`scraped_at`	`string`	ISO-8601 UTC timestamp
`status_code`	`int64`	HTTP response code
`title`	`string`	`<title>` tag text
`description`	`string`	`<meta name=description>` content
`keywords`	`string`	`<meta name=keywords>` content
`author`	`string`	`<meta name=author>` content
`lang`	`string`	`<html lang>` attribute
`canonical`	`string`	Canonical URL from `<link rel=canonical>`
`og_tags`	`string`	JSON dict of Open Graph `<meta property=og:*>` tags
`twitter_tags`	`string`	JSON dict of Twitter card `<meta name=twitter:*>` tags
`schema_org`	`string`	JSON list of `<script type=application/ld+json>` objects
`text`	`string`	Clean body text (scripts/styles stripped)
`text_all`	`string`	Every visible string including nav, footer, etc.
`word_count`	`int64`	Word count of `text`
`token_estimate`	`int64`	Approximate token count (words × 1.35)
`content_hash`	`string`	SHA-256 prefix of `text` for deduplication
`structured`	`string`	JSON with headings, paragraphs, tables, lists, blockquotes
`html`	`string`	Raw HTML source (capped at 12 MB)
`links`	`string`	JSON list of internal `<a href>` URLs
`external_links`	`string`	JSON list of external `<a href>` URLs
`images`	`string`	JSON list of image objects (url, alt, base64 data if fetched…)
`code_blocks`	`string`	JSON list of `<code>/<pre>/<script>/<style>` blocks
`error`	`string?`	`null` on success; error string on failure

💻 Usage Examples

from datasets import load_dataset

# Load a full scrape batch
ds = load_dataset("ray0rf1re/unclean-web", split="batch_20260101_120000")

# Load the 1M-token subset
ds = load_dataset("ray0rf1re/unclean-web", split="subset_1m_tokens")

# Load OpenHermes-format for fine-tuning
ds = load_dataset("ray0rf1re/unclean-web", split="subset_openhermes")

# Load tokenizer training shards
ds = load_dataset("ray0rf1re/unclean-web", split="subset_tokenizer")

# Stream very large splits
ds = load_dataset("ray0rf1re/unclean-web", split="merged", streaming=True)

# Work with structured content
import json
for row in ds:
    structured = json.loads(row["structured"])
    for heading in structured["headings"]:
        print(heading["level"], heading["text"])

# Iterate images (if fetched)
for row in ds:
    for img in json.loads(row["images"]):
        if "data_b64" in img:
            import base64
            raw = base64.b64decode(img["data_b64"])

🌐 Scraped Sources

URL	Domain
`http://blog.exoplatform.org`	blog.exoplatform.org
`http://technovisions.wordpress.com`	technovisions.wordpress.com
`http://www.myheritage.com`	www.myheritage.com
`https://ao3wiki.com`	ao3wiki.com
`https://aoya.moe`	aoya.moe
`https://archive.transformativeworks.org`	archive.transformativeworks.org
`https://archiveofourown.gay`	archiveofourown.gay
`https://archiveofourown.org`	archiveofourown.org
`https://bloggers.feedspot.com`	bloggers.feedspot.com
`https://bloggerspassion.com`	bloggerspassion.com
`https://en.wikipedia.org`	en.wikipedia.org
`https://fanlore.org`	fanlore.org
`https://github.com`	github.com
`https://itch.io`	itch.io
`https://m.web/`	m.web
`https://medium.com`	medium.com
`https://novel.com/`	novel.com
`https://wiki.archiveteam.org`	wiki.archiveteam.org
`https://www.epubor.com`	www.epubor.com
`https://www.fandom.com`	www.fandom.com
`https://www.fanfiction.net`	www.fanfiction.net
`https://www.fanon.co`	www.fanon.co
`https://www.ifarchive.org`	www.ifarchive.org
`https://www.ontoplist.com`	www.ontoplist.com
`https://www.reddit.com`	www.reddit.com
`https://www.tech2geek.net`	www.tech2geek.net
`https://www.webtoons.com`	www.webtoons.com

📅 Update Log

Timestamp (UTC)	Pages	Tokens	Sites	Split
2026-05-21 04:40 UTC	261	256.7K	`https://www.webtoons.com`	`batch_20260521_033424`
2026-05-20 17:40 UTC	11,656	0	`https://www.ifarchive.org`, `https://www.epubor.com`, `https://www.fanon.co`…`	`batch_20260520_134516`
2026-05-20 08:18 UTC	4,201	0	`https://www.reddit.com`, `https://archiveofourown.org`, `https://fanlore.org`…`	`batch_20260520_070044`
2026-05-20 06:59 UTC	7	0	`https://bloggerspassion.com`, `http://technovisions.wordpress.com`, `http://blog.exoplatform.org`…`	`batch_20260520_065855`
2026-05-20 13:33 UTC	2	0	`https://m.web/`, `https://novel.com/`	`batch_20260520_133214`
2026-05-21 00:54 UTC	2,738	0	`https://github.com`	`batch_20260520_223017`

🧰 Pipeline

This dataset is produced by scrape_to_hf.py, a Python pipeline that features:

Smart URL parsing — comma, space, or zero separator between sites
BFS recursive crawl — configurable depth (1–25), same-origin only
Temp-cache per depth — each depth level is cached to disk then flushed, enabling large crawls without OOM
Token budgeting (-M) — stop when a token target is reached
Auto site discovery (-A) — finds sites via DuckDuckGo, Bing, Common Crawl
Retry + backoff — 3 retries with exponential backoff, robots.txt respected
Parallel workers — multi-threaded BFS for faster crawls
Automatic subsets — 7 standard subsets generated after each run
Self-updating README — this card is regenerated on every upload

⚠️ Disclaimers

Content is raw and unfiltered — may contain offensive, adult, or sensitive material.
Image data is stored as base64 in the images JSON column (only when --no-images is not set).
robots.txt is respected by default. Pass --ignore-robots to override.
Each upload run creates a new named split so data accumulates rather than overwriting.

📄 Citation

@misc{unclean-web,
  author       = {ray0rf1re},
  title        = {Unclean Web — Raw Internet Crawl Dataset},
  year         = {2026},
  publisher    = {HuggingFace},
  howpublished = {\url{https://huggingface.co/datasets/ray0rf1re/unclean-web}},
}

Dataset card auto-generated by scrape_to_hf.py v3.0 — 2026-05-21 04:40 UTC

Downloads last month: 120