pauljeffrey/westafrican-mt-rlhf

A distributed fine-tuning framework that applies RLHF with AfriCOMET as a reward model to improve machine translation for low-resource and under-resourced language pairs. # West African Machine Translation with AfriCOMET-Guided RLHF **A machine translation system for West African languages. It teaches a small language model to translate, then improves it using feedback from a quality-scoring model built for African languages. Work is still ongoing.** | | | |---|---| | **Base model** | `google/gemma-3-270m-it` | | **Published checkpoint** | `BeardedMonster/gemma-3-270m-translate-it` | | **Training data** | `Aletheia-ng/tds-sft` | | **Reward model** | `masakhane/africomet-stl` | | **Evaluation benchmark** | `masakhane/mafand` (validation split) | | **Languages** | Hausa, Igbo, Yoruba, Wolof, Ewe, Fon, Twi (+ Nigerian Pidgin in training data) | --- ## Table of Contents 1. Problem Statement 2. Approach 3. How It Works 4. Evaluation 5. Related Academic Work 6. Quick Start 7. Multi-GPU Training 8. Project Layout 9. Tech Stack --- ## 1. Problem Statement Most translation tools work well for languages like English, French, and German — but **West African languages are underserved**. Hausa, Igbo, Yoruba, Wolof, and others have far less training data, and general-purpose AI models often produce weak translations for them. Three challenges make this hard: 1. **Not enough data** — There are far fewer high-quality translation examples for these languages than for major world languages. 2. **Hard to measure quality** — Standard translation scores do not always reflect how good a translation actually sounds to speakers of the language. 3. **Training cost** — Improving a model on millions of examples, then refining it with quality feedback, takes real compute — even with a small model. **What this project does:** - Trains **Gemma 3 270M**, a compact Google language model, on `Aletheia-ng/tds-sft` (~11M translation examples). - Runs a second training stage that **rewards better translations** using **AfriCOMET**, a quality scorer built for African languages. - Supports **single-GPU or multi-GPU training** so the same pipeline works on a l …