README.md

metadata

language:
  - el
license_name: common-crawl-terms-of-use
license_link: https://commoncrawl.org/terms-of-use
pretty_name: Greek Common Crawl
task_categories:
  - text-generation
  - fill-mask
tags:
  - greek
  - common-crawl
  - web-text
  - fineweb
license: odc-by

Greek Common Crawl

A FineWeb-style Greek-language text dataset extracted from Common Crawl, following the FineWeb-2 recipe adapted for Greek (ell_Grek).

Pipeline source: github.com/alexliap/greek-cc.

Dataset structure

Each processed crawl (e.g. CC-MAIN-2024-22) contributes two kinds of files:

• raw/<crawl>/chunk=<offset>/*.parquet: per-chunk survivors, written as soon as each ~100k-row manifest chunk finishes extraction and filtering. These are pre-deduplication: near-duplicate detection only runs once a whole crawl is fully processed, so documents here may have near-duplicates elsewhere in the same crawl.
• <crawl>/*.parquet: the final, whole-crawl output, written once every chunk is done. Every document here has been scored for near-duplicate clusters (see below).

Columns

column	type	description
id	string	stable id (content digest)
text	string	extracted, cleaned document text
crawl	string	source crawl id, e.g. CC-MAIN-2024-22
content_digest	string	Common Crawl content digest
url	string	source page URL
url_host_registered_domain	string	registered domain of url
warc_filename	string	source WARC file in the crawl
warc_record_offset	int64	byte offset of the WARC record
language	string	predicted language (GlotLID)
language_score	float64	GlotLID confidence for language
language_script	string	predicted script
minhash_cluster_id*	int	near-duplicate cluster id
minhash_cluster_size*	int	size of that cluster

* only present in the final <crawl>/*.parquet files, not in raw/.

Near-duplicates are tagged, not removed. Documents that MinHash considers near-duplicates of each other share a minhash_cluster_id, and the final files still contain every document. If you want a deduplicated set, keep one document per minhash_cluster_id yourself (e.g. the longest, or the first seen).

How it was built

1. Manifest: Common Crawl's columnar index is scanned for Greek-flagged records, producing a Parquet manifest of (warc_filename, offset, length) per candidate record.
2. Fetch and extract: each manifest row is range-fetched from the crawl's WARC files on S3 and run through Trafilatura for boilerplate removal.
3. Filter, in order:
   • URL filter (blocklists)
   • Language ID (GlotLID), kept only if predicted language is Greek (ell) above a tuned confidence threshold
   • Gopher repetition filter
   • FineWeb quality filter
   • Gopher quality filter
   • all thresholds are FineWeb-2's own Greek (ell_Grek) config, used as-is rather than hand-tuned
4. Clean: FTFY (mojibake/encoding fixes), PII redaction, stray symbol line removal.
5. Deduplicate, per-snapshot rather than per-language-global. MinHash (14 buckets, 8 hashes per bucket, 5-grams, xxhash 64-bit, the same hashing scheme FineWeb-2 uses), but scoped to one crawl at a time instead of across every crawl. FineWeb-2 dedups Greek globally across its full 2013 to April 2024 span, then has to upsample surviving documents by cluster size to claw back the volume that global dedup throws away. The original FineWeb (v1) found the opposite worked better: per-snapshot dedup kept far more tokens (20T vs 4T in their ablation) and avoided the quality drop global dedup caused, where what survives skews toward ads and keyword-stuffed text. This project follows FineWeb v1's per-snapshot scope for that reason. Documents are also tagged rather than dropped (see above), which is a further departure from both FineWeb v1 and FineWeb-2, who each remove near-duplicates outright. Pre-deduplication data is available in case someone wants to do deduplication globally, the FineWeb 2's way.

Source and license

Licensed under ODC-By 1.0, matching FineWeb-2's ell_Grek. Built from Common Crawl WARC data, so Common Crawl's own Terms of Use also apply to the underlying web content.

References

• Penedo et al., The FineWeb Datasets: Decanting the Web for the Finest Text Data at Scale (arXiv 2406.17557)
• Penedo et al., FineWeb2: One Pipeline to Scale Them All (arXiv 2506.20920)
• datatrove, the processing library both FineWeb papers and this pipeline are built on

Status

Crawls are processed and uploaded incrementally. Check which <crawl>/ directories exist for what's finished, and raw/<crawl>/ for crawls still in progress.