Benjamin-png/swahili-pretraining-corpus

Swahili Pretraining Corpus (tokenized)

The pre-training corpus used to train Benjamin-png/swahili-gpt-71m from scratch — ~2.15 billion tokens of real Swahili text, already tokenized and ready for language-model training.

Contents

File	What it is
train.bin	Training tokens — a flat uint16 array of token IDs (~2.15B tokens, ~4.3 GB)
val.bin	Held-out validation tokens (same format)
swahili_tokenizer.model	The SentencePiece tokenizer (32k vocab, byte_fallback) used to produce the IDs
meta.json	Token counts, vocab size, dtype, and the list of sources

How to read it

import numpy as np, sentencepiece as spm
data = np.memmap("train.bin", dtype=np.uint16, mode="r")   # token IDs
sp = spm.SentencePieceProcessor(); sp.load("swahili_tokenizer.model")
print(sp.decode(data[:50].tolist()))                       # peek at the text

The IDs correspond only to the included tokenizer (vocab 32000). The tokenizer was trained on 30M cased sentences, with byte_fallback=True so it never emits <unk> (handles capitalization, proper nouns, and arbitrary characters).

Sources & attribution

This is a derivative, tokenized blend. Full credit to the original corpora:

• Inkuba-Mono-Swahili (Alfaxad) — the bulk of the data.
• Mendeley Swahili Corpus (CC BY 4.0).
• A locally-cleaned Swahili corpus.

meta.json lists the sources included (sources_done).

Why publish the tokenized form?

Sharing the exact train.bin + tokenizer makes the 71M model's training fully reproducible — anyone can train on the identical token stream. To rebuild from the raw sources instead, see build_corpus_bin.py in the project repo.

License

CC BY-4.0, matching the source corpora. Please keep attribution.