FineWeb-CLaR-region: region-attributed index of the multilingual web

A region-attributed document index over FineWeb and FineWeb-2: 7,916,767,586 documents across 6,332 locales (language–script–region combinations; 346 languages). Each row is a pointer — document id, CommonCrawl dump, URL, and region-attribution metadata — not the text itself. Text is joined back from the origin corpora by document id (code below).

For a ready-to-use annotated subset with text (≤10,000 docs/locale, topic + cultural-taxonomy labels), see Yusser/FineWeb-CLaR-culture.

Layout

Hive-partitioned parquet, one directory per locale:

data/language_script=deu_Latn/region=MD/fineweb2-part-0000.parquet

Columns

column	type	description
id	str	origin document id (<urn:uuid:...>), the join key into FineWeb / FineWeb-2
dump	str	CommonCrawl dump, e.g. CC-MAIN-2022-05
url, date, file_path	str	source URL, crawl date, CC WARC path
language, language_script	str	language code / <lang>_<Script>
language_score	float	origin language-id confidence
token_count, minhash_cluster_size, top_langs		origin corpus metadata (may be null)
region, locale	str	attributed region code and <lang>_<Script>-<REGION>
region_source, region_confidence, region_top_rule	str	which attribution signal fired (e.g. url_hint)
region_content_share	float	content-signal share supporting the region
source_dataset	str	origin corpus: fineweb (English) or fineweb2 (all other languages)

Usage

Load the pointers for a locale

import pyarrow.dataset as pad

d = pad.dataset("hf://datasets/Yusser/FineWeb-CLaR-region/data", partitioning="hive")
ptr = d.to_table(
    columns=["id", "dump", "url", "source_dataset"],
    filter=(pad.field("language_script") == "deu_Latn") & (pad.field("region") == "MD"),
)

Get the source text from FineWeb / FineWeb-2

Rows carry no text; join it back from the origin corpus by document id. The source_dataset column picks the corpus (fineweb for English locales, fineweb2 for everything else), dump narrows FineWeb to the right subset, and language_script names the FineWeb-2 config:

from datasets import load_dataset

ids = set(ptr.column("id").to_pylist())
source = ptr.column("source_dataset")[0].as_py()

if source == "fineweb":                      # English: one config per CC dump
    streams = [
        load_dataset("HuggingFaceFW/fineweb", name=dump, split="train", streaming=True)
        for dump in sorted(set(ptr.column("dump").to_pylist()))
    ]
else:                                        # FineWeb-2: one config per language
    streams = [load_dataset("HuggingFaceFW/fineweb-2", name="deu_Latn",
                            split="train", streaming=True)]

texts = {}
for stream in streams:
    for row in stream:
        if row["id"] in ids:
            texts[row["id"]] = row["text"]
            if len(texts) == len(ids):
                break

Streaming scans the whole config to find the matching ids — fine for small locales, slow for large ones. For bulk extraction, snapshot_download the origin parquet shards of the relevant dumps and filter by id (or restrict FineWeb-2 by its dump column) with pyarrow instead of streaming.

References

@inproceedings{penedo2024fineweb,
  title     = {The FineWeb Datasets: Decanting the Web for the Finest Text Data at Scale},
  author    = {Penedo, Guilherme and Kydl{\'i}{\v{c}}ek, Hynek and Ben Allal, Loubna and Lozhkov, Anton and Mitchell, Margaret and Raffel, Colin and Von Werra, Leandro and Wolf, Thomas},
  booktitle = {NeurIPS Datasets and Benchmarks},
  year      = {2024}
}

@article{penedo2025fineweb2,
  title   = {FineWeb2: One Pipeline to Scale Them All -- Adapting Pre-Training Data Processing to Every Language},
  author  = {Penedo, Guilherme and Kydl{\'i}{\v{c}}ek, Hynek and Sabol{\v{c}}ec, Vinko and Messmer, Bettina and Foroutan, Negar and Jaggi, Martin and von Werra, Leandro and Wolf, Thomas},
  journal = {arXiv preprint arXiv:2506.20920},
  year    = {2025}
}

Provenance & license

Document pointers and metadata derive from FineWeb / FineWeb-2 (ODC-By 1.0, subject to the CommonCrawl ToU) and are redistributed under the same ODC-By 1.0 terms. Region attribution is fully automatic (URL country-code hints and content-based signals) — no human labeling.