734ai/maasai-lang

Modern English ↔ Maasai translation and speech transcription showcase for language preservation, accessibility, and low-resource AI research. # Maasai Language Showcase ## English ↔ Maasai Translation + Maasai Speech Transcription A modern English ↔ Maasai translation and Maasai speech transcription showcase built for language preservation, accessibility, and low-resource AI research. --- ## Features - **English → Maasai translation** — powered by a QLoRA fine-tuned model - **Maasai → English translation** — bidirectional support - **Maasai speech transcription** — via Microsoft Paza ASR - **Speech → Translation pipeline** — transcribe then translate - **Cultural terminology preservation** — 90+ term glossary with protected terms - **Sub-tribe awareness** — covers Ldikiri, Laikipiak, Samburu (Lmomonyot), Ilkisongo, Ilpurko, and more ## Architecture | Component | Model | Purpose | |-----------|-------|---------| | Translation | `Qwen/Qwen2.5-3B-Instruct` (QLoRA) | English ↔ Maasai text translation and Maa composition | | Speech | `microsoft/paza-whisper-large-v3-turbo` | Maasai speech transcription | | UI | Gradio | Interactive Hugging Face Space | ## Quick Start ```bash # Setup python -m venv .venv && source .venv/bin/activate pip install -r requirements.txt cp .env.example .env # Generate training data python scripts/generate_cultural_pairs.py python scripts/prepare_data.py # Train bash training/run_train.sh # Evaluate python scripts/evaluate_mt.py # Run Space locally python space/app.py ``` ## Publish to Hugging Face ```bash # Optional: set token explicitly (recommended) export HF_TOKEN=hf_xxx # Preview publish plan (no uploads) python scripts/publish_to_hf.py # Export refreshed publish bundles locally python scripts/publish_to_hf.py --export-dir dist/hf_publish --create-model-repo # Execute publish for Space + dataset + model repo scaffold python scripts/publish_to_hf.py --execute --create-model-repo ``` The publisher uploads: - Space bundle from `space/` + glossary data - Dataset bundle from `data/final_v3/` + dataset card - Model artifacts from `outputs/maasai-en-mt-qlora` when real …