SsemuliJoseph/luganda-nlp-pipeline

# Luganda NLP Pipeline An end-to-end English ↔ Luganda machine translation pipeline: data annotation, baseline fine-tuning, QLoRA adaptation, evaluation, and a deployed demo — built during an industrial training placement at Sunbird AI. ## Overview This project explores low-resource NLP for Luganda, a Bantu language spoken in Uganda. It covers: - **Annotation pipeline** — golden evaluation set curated via Argilla, reviewed under an Accept / Needs Editing / Reject discipline - **Baseline model** — fine-tuned NLLB-200 (distilled 600M) for direct English→Luganda translation - **QLoRA adaptation** — Qwen2.5-1.5B-Instruct fine-tuned with a LoRA adapter in instruction/chat format - **Evaluation** — BLEU scoring plus qualitative side-by-side comparison between baseline and LoRA outputs on the same held-out set - **Deployment** — a live Gradio demo on Hugging Face Spaces ## Results | Model | BLEU | |---|---| | NLLB baseline | 17.84 | ## Links - 🤗 Baseline model: SsemuliJoseph/nllb-luganda-baseline - 🤗 LoRA adapter: SsemuliJoseph/qwen2.5-1.5b-luganda-lora - 🚀 Live demo: luganda-translator-demo Space ## Pipeline Structure The notebook (`Luganda_NLP_Pipeline.ipynb`) is organized into phases: 1. **Phase 0** — Environment setup, authentication, sanity checks 2. **Phase 1** — Annotation pull, golden dataset construction 3. **Phase 2** — Baseline model training and qualitative evaluation 4. **Phase 3** — Tokenizer inspection, push baseline to Hub 5. **Phase 4** — QLoRA fine-tuning, adapter push, comparison vs. baseline 6. **Phase 5** — Robustness utilities (OOM handling, gradient checks) 7. **Phase 6** — Gradio demo, Hugging Face Spaces deployment 8. **Phase 7** — Closing checklist ## Setup \`\`\`bash git clone github.com cd luganda-nlp-pipeline pip install -r requirements.txt \`\`\` ## Author Joseph Ssemuli — Computer Science, Mbarara University of Science and Technology (MUST). This work was developed during an indust …