henryhale/luganda-pretrain

Luganda Pretraining Corpus

A cleaned, normalized, and deduplicated pretraining corpus for the Luganda language (lg), ready to train language models or tokenizers.

Dataset Statistics

• Total Documents: 83,347
• Total Size: 203.7 MB (clean text)
• Primary Format: Parquet (text, source columns)

Data Sources & Composition

This corpus was built by pulling from diverse web and curated datasets, ordered by clean-source priority during deduplication:

1. Wikipedia (20231101.lg): 3,139 docs
2. FineWeb-2 (lug_Latn): 35,108 docs
3. HPLT 2.0 (lug_Latn): 8,371 docs
4. MADLAD-400 Clean: 12,520 docs
5. GlotCC-V1: 714 docs
6. CC-100: 16,935 docs
7. MADLAD-400 Noisy: 6,560 docs

Curation & Quality Filtering Pipeline

• Normalization: Unicode NFC normalization and blank line stripping.
• Quality Heuristics:
  • Retained documents with length >= 30 characters (MIN_CHARS = 30).
  • Required >= 60% of characters to be alphabetic (MIN_ALPHA = 0.6) to filter out tables, code, and noise.
  • Retained documents with <= 30% line-level duplicate boilerplate (MAX_DUP_LINES = 0.3).
• Language ID: Filtered using GlotLID (__label__lug_Latn) with a minimum confidence threshold of >= 0.5 (LID_THRESHOLD = 0.5).
• Exact Deduplication: Removed identical documents using MD5 hashing.
• Near-Deduplication: Applied MinHash LSH using word 5-grams with a Jaccard similarity threshold >= 0.8 (NEAR_DUP = 0.8).

How to Use

from datasets import load_dataset

dataset = load_dataset("henryhale/luganda-pretrain")
print(dataset["train"][0])