MichaelR207/high-quality-cc-21b

high_quality

A high-quality English web text corpus extracted from Common Crawl WARC files using an LLM-based extraction and quality pipeline.

Dataset Summary

high_quality is a pretraining-grade corpus of cleaned web documents. Raw Common Crawl WARC records are passed through an LLM-based extractor that strips boilerplate and recovers the main content, then filtered to retain only documents in the "high_quality" band, deduplicated (exact + fuzzy), and decontaminated against common evaluation benchmarks.

• Roughly 21 billion tokens.
• Sourced from 10,364 Common Crawl WARC files drawn from the DCLM 400m-1x pool.
• Distributed as parquet shards.
• Every row carries full provenance back to its original Common Crawl WARC record.

Exact post-join counts:

• Documents: 19,968,996
• Tokens: ~23.4 billion (estimated as characters / 4; ~21B by tokenizer count)
• Shards: 512 parquet files
• On-disk size: 59.18 GB

Provenance & Traceability

Each document retains the metadata needed to trace it back to the exact Common Crawl record it was extracted from. This makes the corpus auditable: you can re-fetch the original raw record, verify the extraction, or recover additional context (HTTP headers, original HTML).

Per-row provenance fields:

• url — the source page URL.
• warc_file — the Common Crawl WARC file path the document came from.
• warc_record_id — the WARC record UUID within that file.
• snapshot — the Common Crawl crawl snapshot, in CC-MAIN-YYYY-WW form.

To fetch an original record, locate warc_file within the corresponding snapshot on the Common Crawl public data, then seek to the record identified by warc_record_id. The url field lets you cross-check the recovered record against the page it was crawled from.

Coverage. 99.99% of documents (19,967,190 of 19,968,996) carry full WARC provenance. A small remainder — 1,806 documents (0.009%) — retain their text but have null provenance fields: their exact extracted text could not be matched back to a source record during the metadata join. They are kept for completeness; filter on a non-null warc_file if you need only fully-traceable rows.

How it was built

The pipeline runs the following stages:

1. LLM-based extraction / cleaning. Each raw Common Crawl WARC record is processed by an LLM-based extractor that removes navigation, boilerplate, and markup, and reconstructs the readable main text of the page.
2. Quality filtering. Extracted documents are scored and only those in the "high_quality" band are kept.
3. Deduplication. Exact duplicate removal followed by fuzzy near-duplicate removal using MinHash-LSH, to collapse documents that are byte-identical or close paraphrases/near-copies.
4. Decontamination (CORE-v2). Documents containing distinctive evaluation n-grams are dropped so the corpus is safe to train on without leaking common benchmarks (see below).

Decontamination note

The corpus is decontaminated using CORE-v2: documents that contain distinctive n-grams from common evaluation benchmarks are removed. The goal is to let downstream models train on this data without contaminating those benchmarks. Decontamination reduces but does not provably eliminate all overlap — n-gram filtering is a heuristic, and novel or reformatted eval items may still slip through. Treat it as a strong best-effort safeguard, not a guarantee.

Data fields

Field	Type	Description
text	str	The cleaned document text (LLM-extracted main content).
url	str	Source page URL.
warc_record_id	str	WARC record UUID within the source WARC file.
warc_file	str	Common Crawl WARC file path the document was extracted from.
snapshot	str	Common Crawl crawl snapshot (CC-MAIN-YYYY-WW).

Splits & size

The dataset is released as a single train split, sharded into parquet files.

Split	Documents	Tokens	Shards	Size
train	19,968,996	~23.4B	512	59.18 GB

Licensing

The dataset compilation is released under the Open Data Commons Attribution License (ODC-BY 1.0) — the same license used by comparable Common Crawl–derived corpora such as C4 and FineWeb. ODC-BY covers the curation and packaging of this dataset; it does not grant any rights over the underlying web content, which Common Crawl itself does not own.

Common Crawl is not distributed under a Creative Commons license. It is governed by the Common Crawl Terms of Use, and the crawled pages remain the property of their original authors. Accordingly, use of the documents in this dataset is also subject to the Common Crawl Terms of Use, and individual documents remain subject to the rights of their original publishers.

Disclaimer. This data is a derivative of publicly crawled web pages. Use is subject to the Common Crawl Terms of Use, and downstream users are expected to respect robots.txt and the Common Crawl ToU. The underlying content was authored by third parties; the dataset maintainers do not claim ownership of the source text and make no representation about the rights status of any individual document. If you are a content owner and want a document removed, contact the maintainers.

Limitations & biases

• LLM-extraction artifacts. The main text is reconstructed by an LLM, which can introduce errors: dropped or duplicated passages, hallucinated connective text, lost tables/formatting, or occasional misreading of page structure.
• Web noise. Despite quality filtering, the corpus reflects the open web: it contains factual errors, spam-adjacent content, dated information, and the demographic, topical, and linguistic skews of crawled English-language pages.
• English only. The corpus is filtered to English; it is not suitable as a multilingual resource.
• Quality filtering is a heuristic. The "high_quality" band is a model-scored judgment, not ground truth; some low-value documents are retained and some good documents are dropped.
• Decontamination is best-effort (see the decontamination note above).

Citation

@misc{high_quality_cc_21b,
  title  = {high_quality: An LLM-extracted high-quality web text corpus from Common Crawl},
  author = {MichaelR207},
  year   = {2026},
  note   = {Derived from Common Crawl.}
}