AvelonLabs/cc-main-2026-04-cleaned-13k

CC-MAIN-2026-04 Cleaned

Cleaned and filtered text corpus extracted from the Common Crawl January 2026 snapshot (CC-MAIN-2026-04), processed from raw WET files into Parquet format.

Intended for language model pretraining and finetuning.

Dataset Details

Field	Value
Source snapshot	CC-MAIN-2026-04 (January 2026)
Source format	WET (plain text extractions)
Output format	Parquet
Documents kept	~13,847 per segment
Filter drop rate	~40%
Processing time	~4–5 min per segment on T4 GPU

Data Fields

Column	Type	Description
url	string	Source URL of the document
text	string	Cleaned plain text content

Filtering Pipeline

Documents are dropped if they match any of the following:

• Junk URL — adult/porn/escort/cam domains, casino/gambling domains, pharma spam, payday loans, warez, .onion
• Low quality — alpha ratio < 50%, fewer than 3 real sentences, keyword stuffing (single word > 15% of tokens), >50% uppercase, outside 200–50,000 char range
• Junk text — SEO spam, "buy now" / "click here" copy, crack/serial download pages, guaranteed investment scams, casino bonus copy, pharma discount offers, boilerplate walls
• Adult content — 50+ keyword/phrase matches covering explicit terms, platforms, and services
• Gambling — 30+ keyword/phrase matches covering casino platforms, game types, and bonus copy
• Profanity / slurs — ~40 terms matched on full word boundaries

Drop Stats (Segment 0)

Total processed : 23,073
Kept            : 13,847  (60%)
Dropped         :  9,226  (40%)

  low_quality  →  6,678
  junk_url     →    748
  junk_text    →    663
  adult        →    492
  bad_words    →    388
  gambling     →    257

Usage

from datasets import load_dataset

ds = load_dataset("Tralalabs/cc-2026-04-clean")
print(ds["train"][0])

Source Data

Raw data sourced from Common Crawl under their Terms of Use. Common Crawl data is released under CC0 1.0.