README.md

metadata

language:
  - af
  - en
  - nr
  - nso
  - st
  - ss
  - tn
  - ts
  - ve
  - xh
  - zu
tags:
  - pretraining
  - south-african-languages
  - multilingual
  - mzansitext
  - deduplicated
license: apache-2.0
dataset_info:
  features:
    - name: text
      dtype: string
    - name: lang
      dtype: string
  splits:
    - name: validation
      num_bytes: 50448243
      num_examples: 19940
    - name: test
      num_bytes: 50257814
      num_examples: 19784
    - name: train
      num_bytes: 15178461663
      num_examples: 3744654
  download_size: 9577076320
  dataset_size: 15279167720
configs:
  - config_name: default
    data_files:
      - split: validation
        path: data/validation-*
      - split: test
        path: data/test-*
      - split: train
        path: data/train-*

MzansiText — deduplicated release

This is a conservatively deduplicated release of MzansiText, a multilingual pretraining corpus for all eleven official South African languages.

Use the original release to reproduce the paper and its trained models. Use this release for new experiments where cross-source duplicate documents should be removed.

Dataset details

• Splits: 3,744,654 train rows, 19,940 validation rows, and 19,784 test rows
• Total: 3,784,378 rows
• lang values: afr, eng, nbl, nso, sot, ssw, tsn, tso, ven, xho, zul
• Schema: text (string), lang (string)
• All eleven languages occur in every split.

Validation and test sets are capped at approximately 2M tokens per language to prevent high-resource languages from dominating early stopping.

Deduplication

Deduplication is performed before deterministic splitting. Each document is normalized with Unicode NFC, whitespace is collapsed and trimmed, and exact duplicates are identified by (lang, SHA-256(normalized text)). The language is part of the key so identical text labelled as different languages is retained.

The process removed 197,926 duplicate rows:

Language label	Rows removed
afr	165,623
zul	19,156
xho	12,910
tso	231
sot	3
eng	1
nbl	1
nso	1
Total	197,926

An audit of the combined train, validation, and test splits found no remaining duplicates under this normalization and per-language key.

Usage

from datasets import load_dataset

dataset = load_dataset("uctnlp/mzansi-text-deduplicated")
print(dataset)

Reproducibility

• Original paper-reproduction dataset: uctnlp/mzansi-text
• Deduplication implementation and config: Anri-Lombard/sallm#123
• Paper: arXiv:2603.20732

Citation

@misc{lombard2026mzansitextmzansilmopencorpus,
      title={MzansiText and MzansiLM: An Open Corpus and Decoder-Only Language Model for South African Languages},
      author={Anri Lombard and Simbarashe Mawere and Temi Aina and Ethan Wolff and Sbonelo Gumede and Elan Novick and Francois Meyer and Jan Buys},
      year={2026},
      eprint={2603.20732},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2603.20732},
}

License

Apache License 2.0