Reproducible LLM training framework for South African languages. Contains code for MzansiLM (LREC 2026).
# MzansiText & MzansiLM
**An open corpus and decoder-only language model for South African languages.**
This repository accompanies **"MzansiText and MzansiLM: An Open Corpus and
Decoder-Only Language Model for South African Languages"**
(arXiv:2603.20732).
The public workflow is recipe-first: choose a recipe ID, inspect the Hydra
configs it resolves to, then launch fine-tuning or evaluation.
```bash
uv sync
uv run sallm recipes list
uv run sallm recipe show llama_t2x_xho
uv run sallm finetune llama_t2x_xho --dry-run
uv run sallm evaluate llama_t2x_xho --dry-run
```
Recipe IDs are defined in `recipes/registry.yaml`. Each recipe points to known
Hydra configs under `src/conf`; those YAML files remain the source of truth for
hyperparameters.
## Docs
- Quickstart: install, inspect recipes, and dry-run the CLI.
- Recipes: supported recipe IDs and their config targets.
- Configuration: how recipe targets map to Hydra YAML.
- SLURM: advanced cluster script compatibility notes.
## Paper Snapshot
For exact reproduction of the LREC 2026 paper results, use the permanent
snapshot:
```bash
git clone
github.com
cd sallm
git checkout tags/mzansitext-mzansilm-lrec2026-v1
```
The `main` branch is actively maintained and may differ from the paper snapshot.
## Releases
- Paper: arXiv:2603.20732
- Model: anrilombard/mzansilm-125m
- Raw corpus: anrilombard/mzansi-text
- Tokenized corpus: anrilombard/mzansi-text-tokenized
- Collection: MzansiLM
## Repository Layout
- `src/main/sallm`: Python package for training, fine-tuning, and evaluation.
- `src/conf`: Hydra configs for model, data, fine-tuning, evaluation, and HPO.
- `recipes/registry.yaml`: public recipe IDs mapped to known configs.
- `ops/slurm`: advanced SLURM workflows for cluster execution.
- `scripts`: local Python utilities.
- `tokenizer`: tokenizer training and processing utilities.
- `data`: dataset preparation and cleaning utilities.
## Development
```bash
uv sync --extra d …