🇭🇹KreyòlAI

English → Haitian Creole · Domain-aware translation for humanitarian and healthcare work

Healthcare — 250-sentence held-out human-translated medical test set (CMU NESPOLE corpus), English → Haitian Creole:

Model BLEU chrF TER (lower is better)
Google Translate 30.87 58.44 47.90
Base NLLB-200 30.03 56.46 50.56
KreyolAI Healthcare v3 29.87 56.94 49.44
Education — results differ by test set, shown here in full rather than picking the better-looking one:
Test set Model BLEU chrF
--- --- --- ---
MIT-Haiti (human-translated) Base NLLB-200 25.87 48.28
MIT-Haiti (human-translated) KreyolAI Education v1 28.63 50.90
v2_education (mined, not human-verified) Base NLLB-200 31.81 56.49
v2_education (mined, not human-verified) KreyolAI Education v1 30.93 56.74
On the human-translated set, the adapter clearly improves on the base model (+2.76 BLEU; TER not measured in that earlier run). On the mined test set, results were roughly tied with base (-0.88 BLEU, +1.80 TER) — a reminder that scores can shift with the test set used.
Humanitarian — mined test set, not independently human-translated; treat as a preliminary result:
Model BLEU chrF TER (lower is better)
--- --- --- ---
Base NLLB-200 28.26 54.54 55.53
KreyolAI Humanitarian v1 29.23 55.87 54.05
A modest +0.98 BLEU edge over base, and a lower (better) TER by 1.48 points too. A rigorous human-translated evaluation for this domain is still a priority.
Honest interpretation: across all three domains, these models are not a benchmark blowout over the best available systems — they range from roughly tied to modestly ahead of the base model, depending on the test set. The real value isn't a higher score, it's a free, offline, domain-specialized translator for a language most providers treat as an afterthought.
Google Translate score obtained via the public translation endpoint, for reference only — not an official Google benchmark. Full details on each model's page: huggingface.co/dondodoai
📝 English Input
Domain

Auto detects from your text

🌐 Kreyòl Output

If the translation is wrong, submit the correct version. Every correction helps improve the model.

💡 Example Translations
Examples
Domain