SaarAI/fineweb-2-malagasy

FineWeb-2 Malagasy

The Malagasy subsets of HuggingFaceFW/fineweb-2 in one repository, plus long Malagasy Wikipedia articles, deduplicated across all of them. FineWeb-2 is a deduplicated and filtered Common Crawl web corpus for pretraining.

FineWeb-2 has no mlg / mg config. Malagasy is split by GlotLID into three ISO 639-3 varieties:

config	variety	description
plt_Latn	Plateau Malagasy	Standard Malagasy (the official language of Madagascar)
skg_Latn	Sakalava Malagasy	Western-coast dialect
xmv_Latn	Antankarana Malagasy	Northern dialect

from datasets import load_dataset

ds = load_dataset("SaarAI/fineweb-2-malagasy")                    # default config: plt_Latn
ds = load_dataset("SaarAI/fineweb-2-malagasy", "skg_Latn")        # Sakalava
ds = load_dataset("SaarAI/fineweb-2-malagasy", "plt_Latn", split="train", streaming=True)

Splits

config	variety	split	documents	words	characters	parquet MB	avg LID score
plt_Latn	Plateau Malagasy	train	272,060	136,270,400	946,551,115	377.69	0.995
plt_Latn	Plateau Malagasy	test	2,445	1,233,542	8,578,009	3.50	0.994
skg_Latn	Sakalava Malagasy	train	808	312,585	1,963,668	0.79	0.986
skg_Latn	Sakalava Malagasy	test	13	3,935	25,222	0.02	0.946
xmv_Latn	Antankarana Malagasy	train	154	165,488	1,195,134	0.31	0.728
xmv_Latn	Antankarana Malagasy	test	1	680	4,054	0.01	0.649
total			275,481	137,986,630	958,317,202	382.32

• train / test are the original FineWeb-2 splits (not re-split), with cross-config duplicates removed (see Deduplication).
• Words = whitespace-separated tokens. LID score = GlotLID confidence for the assigned label.
• plt_Latn is almost all of the data. skg_Latn and xmv_Latn are very small. xmv_Latn has a low average LID score: many of its documents also score highly for plt_Latn (see top_langs).
• The *_removed folders of FineWeb-2 (documents rejected by its filters) are not included.
• Common Crawl dumps covered: 96 (CC-MAIN-2013-20 to CC-MAIN-2024-18).

Columns

Same as FineWeb-2:

column	description
text	Document text.
id	Common Crawl WARC record id.
dump	Common Crawl dump, e.g. CC-MAIN-2021-31.
url, date	Source URL and crawl date.
file_path	Source WARC file on S3.
language, language_script	GlotLID label (plt / skg / xmv) and script (Latn).
language_score	GlotLID confidence for language.
top_langs	JSON string with the top GlotLID label scores.
minhash_cluster_size	Size of the document's MinHash duplicate cluster before deduplication.
wordlist_ratio	FineWeb-2 filtering feature. Only in xmv_Latn (as in the source files).

Malagasy Wikipedia (wikipedia_mg)

Long articles from wikimedia/wikipedia config 20231101.mg (Wikipedia dump of 2023-11-01). Only articles with more than 500 words (whitespace-separated) are kept: 1,811 of 96,316 articles (1.9%). Most Malagasy Wikipedia pages are short bot-made stubs (place names, postcodes, one-line bios); the filter removes them. After deduplication (see below) 1,808 articles with 2,184,649 of 8,907,908 source words (24.5%) remain.

config	split	articles	words	characters	median words	parquet MB
wikipedia_mg	train	1,808	2,184,649	14,867,274	758	5.36

ds = load_dataset("SaarAI/fineweb-2-malagasy", "wikipedia_mg", split="train")

Columns: id, url, title, text (as in the source) plus word_count (the whitespace word count used for the filter). This config has different columns from the FineWeb-2 configs, so it is a separate config. Some overlap with FineWeb-2 is possible, as Common Crawl also crawls mg.wikipedia.org.

Deduplication

FineWeb-2 deduplicates each language on its own. Here all configs and splits are deduplicated together (across plt / skg / xmv, train / test and Wikipedia), removing 818 documents (0.29%):

1. Exact: MD5 of the NFKC-normalized, lowercased, whitespace-collapsed text.
2. Near: MinHash over word 5-gram shingles (112 hashes, LSH with 14 bands x 8 rows). Candidate pairs are merged when their estimated Jaccard similarity is >= 0.8; duplicate clusters are formed with union-find.

From each duplicate cluster one document is kept, chosen by priority wikipedia_mg/train > plt_Latn/test > skg_Latn/test > xmv_Latn/test > plt_Latn/train > skg_Latn/train > xmv_Latn/train, then by length (longest kept). So the test splits are never reduced in favour of train, and train contains no near-copy of a test document.

config/split	before	exact dups	near dups	after
wikipedia_mg/train	1,811	0	3	1,808
plt_Latn/test	2,447	0	2	2,445
skg_Latn/test	13	0	0	13
xmv_Latn/test	1	0	0	1
plt_Latn/train	272,871	4	807	272,060
skg_Latn/train	810	0	2	808
xmv_Latn/train	154	0	0	154
total	278,107			277,289

• 529 removed from plt_Latn/train, copy kept in plt_Latn/train
• 276 removed from plt_Latn/train, copy kept in wikipedia_mg/train
• 6 removed from plt_Latn/train, copy kept in plt_Latn/test
• 3 removed from wikipedia_mg/train, copy kept in wikipedia_mg/train
• 2 removed from plt_Latn/test, copy kept in wikipedia_mg/train
• 2 removed from skg_Latn/train, copy kept in skg_Latn/train

Most removed documents are repeated news / blog pages (e.g. Global Voices archive and category pages) and Common Crawl copies of mg.wikipedia.org articles that are already in wikipedia_mg. Statistics are in dedup_stats.json.

License and citation

• FineWeb-2 configs: same license as FineWeb-2, ODC-By 1.0, also subject to the Common Crawl terms of use.
• wikipedia_mg: Wikipedia text is licensed CC BY-SA 4.0 (and GFDL); see Wikimedia terms.

All credit for the FineWeb-2 data goes to its authors:

@misc{penedo2025fineweb2pipelinescale,
      title={FineWeb2: One Pipeline to Scale Them All -- Adapting Pre-Training Data Processing to Every Language},
      author={Guilherme Penedo and Hynek Kydlíček and Vinko Sabolčec and Bettina Messmer and Negar Foroutan and Amir Hossein Kargaran and Colin Raffel and Martin Jaggi and Leandro Von Werra and Thomas Wolf},
      year={2025},
      eprint={2506.20920},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2506.20920},
}

Reproducing

Built with push_fineweb2_malagasy.py (--wiki-min-words 500). Deduplication is in dedup_fineweb2_malagasy.py; apart from removed rows the FineWeb-2 data is unchanged.