Yusser/Fineweb-LARD

FineWeb + FineWeb-2 with served-from region annotations (URL-only, resolved-only)

A region-annotated, hive-partitioned view of HuggingFaceFW/fineweb (the sample-100BT slice) and HuggingFaceFW/fineweb-2. Every record carries an ISO 3166-1 alpha-2 region label produced by the URL-only resolution pipeline described in the companion paper. Records with region == "XX" (unresolved) have been dropped; to access them, use the original FineWeb / FineWeb-2 datasets and rejoin on (source_dataset, id).

• Records: 2,960,753,780 (resolved-only)
• Locales (language_script × region): 30,012
• Languages (resolved): 525 • Regions: 243
• Layout: hive-partitioned data/language_script={ls}/region={r}/{source_dataset}-*.parquet

Resolution rate, not precision. The 57.20 % combined resolution rate (against the full input including XX) is recall-like — a measure of whether any label was produced, not whether it is correct. Label precision is not yet evaluated. Use the region_confidence column to filter (high/medium/low are the lookup-backed tiers; ccTLD and url_hint are weaker; rows with no signal were dropped).

Schema

Column	Type	Notes
text	string	Document body, verbatim from the source dataset.
id	string	Source record id (UUID-like). Together with source_dataset uniquely identifies the source row.
dump	string	CommonCrawl dump id (e.g. CC-MAIN-2017-04).
url	string	The URL used for region resolution.
date	string	WARC date.
file_path	string	Source WARC S3 path (for traceability).
language	string	ISO 639-3 (FineWeb's en is normalised to eng).
language_script	string	ISO 15924 (Latn, Cyrl, Hani, Jpan, Arab, …).
language_score	float64	Source-dataset language-ID confidence.
token_count	int64	FineWeb only (null for FineWeb-2).
minhash_cluster_size	int64	FineWeb-2 only (null for FineWeb).
top_langs	string	FineWeb-2 only (per-segment lang JSON, null for FineWeb).
region	string	ISO 3166-1 alpha-2. XX rows are excluded.
locale	string	<language>_<script>-<region> (e.g. eng_Latn-GB).
region_source	string	query / host_path / host / domain / url_hint / tld.
region_confidence	string	high / medium / low / url_hint / ccTLD.
region_content_share	float32	A11 content-scan vote share (0.0 in this URL-only release).
region_top_rule	string	A1–A8 when the lookup CSV path was consulted, else empty.
source_dataset	string	"fineweb" or "fineweb2".

Access patterns

The repository is hive-partitioned so you can scope a load to a locale, a region, or a language without scanning the whole corpus.

By locale (language_script-region)

from datasets import load_dataset

ds = load_dataset(
    "REPO_ID",
    data_files="data/language_script=eng_Latn/region=GB/*.parquet",
    split="train",
)
print(ds[0])

By region (across all languages)

ds = load_dataset(
    "REPO_ID",
    data_files="data/**/region=DE/*.parquet",
    split="train",
)

By language_script (across all regions)

ds = load_dataset(
    "REPO_ID",
    data_files="data/language_script=rus_Cyrl/**/*.parquet",
    split="train",
)

Named configs (pre-declared in the dataset card)

For the most common slices we predeclare configs so you can do:

ds = load_dataset("REPO_ID", "lang_eng_Latn",  split="train")
ds = load_dataset("REPO_ID", "region_GB",      split="train")
ds = load_dataset("REPO_ID", "fineweb2_only",  split="train")
ds = load_dataset("REPO_ID", "default",        split="train")  # everything

The named configs are convenience shortcuts; the data_files glob route works for any (language_script, region) combination, including those without a named config.

Filter by confidence

region_confidence follows the pipeline's ladder. To get only the lookup-backed tiers (no bare-ccTLD fallback):

ds = ds.filter(lambda x: x["region_confidence"] in ("high", "medium", "low"))

Rejoin to the original corpora

To get the XX (unresolved) rows back, or to verify a row against the source dataset:

# source_dataset ∈ {"fineweb", "fineweb2"}, id is the per-source record id.
src = "HuggingFaceFW/fineweb" if row["source_dataset"] == "fineweb" else "HuggingFaceFW/fineweb-2"
# Streaming lookup; the source datasets do not have an id index, but
# the (dump, url) tuple is unique in practice.

Provenance

• FineWeb rows come from the sample-100BT subset.
• FineWeb-2 rows come from the full ingest (5.03 B records), minus one corrupt parquet shard (lit_Latn-XX/data-00084.parquet, 1.98 M records — all unresolved, so they would have been dropped anyway).
• Region labels are produced by the region_lookup_v3 resolver and its 1.33 M-key URL lookup table. The pipeline is described in the companion paper (paper.md in the source tree).
• A11 content-scan is not enabled in this release. region_content_share is 0 for every row; the region_source = "content" source value never fires.

Limitations

• Precision unmeasured. No held-out gold set has been built yet. Treat labels as best-effort, not certified.
• cmn_Hani is preserved verbatim from FineWeb-2's source directories. Simplified vs Traditional (Hans/Hant) is not distinguished; downstream consumers needing the split must derive it from region (CN/SG → Hans; TW/HK/MO → Hant) or by re-scanning text.
• Non-Latin / RTL URL handling is unaudited at scale. Punycode IDN hosts (xn--*) resolve via the ccTLD fallback but not via host-key lookups built on Latin forms.
• No cross-corpus dedup. A URL that appears in both FineWeb and FineWeb-2 produces two rows.

Citation

@misc{fineweb-region-urlonly-2026,
  title  = {{FineWeb + FineWeb-2 URL-only region annotations}},
  author = {Yusser Al-Ghussin},
  year   = {2026},
  url    = {https://huggingface.co/datasets/REPO_ID},
}

License

Inherits from the source datasets:

• FineWeb: ODC-By 1.0
• FineWeb-2: ODC-By 1.0

The annotation columns added by this work (region, locale, region_source, region_confidence, region_content_share, region_top_rule, source_dataset) are released under CC0-1.0.