atlasia-ma/darija-translator

# Darija Translator Fine-tuning LFM2.5-230M for English↔Darija translation with LoRA (Unsloth + TRL). ## Architecture Pipeline-stage layout, not layered/hexagonal — this is a training pipeline, not a domain-rich application, so the structure follows the stages data moves through rather than DDD-style layers: src/darija_translator/ ├── config.py # DataConfig, ModelConfig, TrainConfig — frozen dataclasses ├── data.py # filtering, chat formatting, length filtering, split ├── model.py # load base model + attach LoRA (Unsloth) ├── train.py # SFTTrainer wiring, wandb tracking ├── evaluate.py # BLEU/chrF scoring + generation └── cli.py #`darija-translator train` / `evaluate` Pure logic (`data.py`, `evaluate.py`'s `compute_translation_metrics`, all of `config.py`) is unit-tested. Model/training code (`model.py`, `train.py`, `generate_translations`) isn't — it's an integration point with a real model and GPU, verified instead via manual smoke-test scripts in `scripts/`. ## Setup uv sync --group dev --group eval # run the test suite, no GPU needed uv run pytest -v For training/evaluation on GPU, also install: uv sync --group dev --group eval --group train ### W&B tracking Training reports to Weights & Biases by default. cp .env.example .env # fill in WANDB_API_KEY from wandb.ai ## Hugging face checkpointing huggingface-cli login ## Training (requires GPU) uv run darija-translator train ## Evaluation uv run darija-translator evaluate Reports BLEU and chrF on the held-out split. ## Results ## Original exploration The initial SFT experiment (before this repo existed) is kept at `notebooks/01_exploration.ipynb` for reference.