henryhale/luganda-pretrain2

Luganda Pretraining Corpus v2

henryhale/luganda-pretrain, plus the Goldfish lug sources and Luganda translations, deduplicated into one corpus.

Columns: text (one source row, lines joined with \n), source.

Sources

Listed in deduplication priority: a line already kept from an earlier source is dropped from later ones.

source	Origin	Rows	Text	Tokens	License
luganda-pretrain	henryhale/luganda-pretrain	80,905	172.2 MB	52.8M	CC BY 4.0
ebible	eBible lug-lug (Biblica, 2014)	30,792	4.1 MB	1.1M	CC BY-SA 4.0
makerere-parallel	English-Luganda parallel corpus, Luganda side	14,997	0.9 MB	0.2M	CC BY 4.0
english-luganda	henryhale/english-luganda-translation, luganda	56,061	5.9 MB	1.7M	CC BY 4.0
glot500	cis-lmu/Glot500, lug_Latn	359,075	45.6 MB	13.3M	Per sub-source, see Glot500
nllb	allenai/nllb, eng_Latn-lug_Latn, Luganda side	869,933	47.4 MB	16.9M	ODC-BY 1.0
tinystories-ganda	henryhale/tinystories-ganda, tgt	11,065,558	814.0 MB	186.9M	CC BY 4.0 (from TinyStories, CDLA-Sharing-1.0)
Total		12,477,321	1.09 GB	273.0M

TinyStories accounts for 68.5% of the tokens. Filter on source to rebalance the mix or to drop a license you can't use.

Processing

• Unicode NFC, ’ → ' (Luganda writes the same elision both ways), whitespace stripped.
• Exact line-level deduplication across all sources, in the priority order above. Empty lines are dropped.
• No other filtering: the upstream datasets are already language-identified and cleaned.

Usage

from datasets import load_dataset

ds = load_dataset("henryhale/luganda-pretrain2", split="train")
web_only = ds.filter(lambda s: s != "tinystories-ganda", input_columns="source")