KreyòlAI — English → Haitian Creole Translation System
Domain-specialized neural machine translation for Haiti. Built on NLLB-200 1.3B with swappable LoRA adapters.
Models on HuggingFace
| Adapter | Domain | Model ID |
|---|---|---|
| Base | General | facebook/nllb-200-distilled-1.3B |
| v7 | General fine-tuned | dondodoai/nllb-200-1.3b-en-ht-v7 |
| Healthcare | Medical / Clinical | dondodoai/nllb-kreyolai-healthcare-v1 |
| Humanitarian | Aid / Crisis | dondodoai/nllb-kreyolai-humanitarian-v1 |
| Education | Educational | dondodoai/nllb-kreyolai-education-v1 |
Architecture
[ English Input Text ]
│
▼
┌─────────────────────────┐
│ Domain Classifier │
│ (keyword routing) │
└────────────┬────────────┘
│
┌──────────────┼──────────────┐
▼ ▼ ▼
Healthcare Humanitarian Education
LoRA Adapter LoRA Adapter LoRA Adapter
└──────────────┼──────────────┘
▼
Base NLLB-200 1.3B
(Frozen weights)
│
▼
[ Haitian Creole Output ]
Evaluation Results
| Model | BLEU | Dataset | Notes |
|---|---|---|---|
| Base NLLB-1.3B | 23.78 | FLORES-200 | Baseline |
| v7 general | 21.34 | FLORES-200 | Fine-tuned |
| Healthcare v1 | 67.33 | In-domain val | Inflated (back-translated eval) |
| Humanitarian v1 | 43.38 | In-domain val | |
| Education v1 | ~28.82 | MIT-Haiti human | Most trustworthy eval |
Use Cases
- Community health worker communication
- Humanitarian aid coordination
- Educational content translation
- NGO field operations in Haiti
Data Sources
- MIT-Haiti corpus (human-translated educational pairs)
- jhu-clsp/kreyol-mt (FR↔HT parallel data)
- CohereLabs/aya_collection_language_split (monolingual HT, back-translated)
- v7 mined EN↔HT data
Built by
James Dominique — AI/ML engineer, Haiti HuggingFace: dondodoai