We're six months into training a large language model for Mauritian Creole (Kreol Morisien). Here's where things stand.
Why Kreol
Kreol is spoken by over 1.3 million people in Mauritius and the diaspora. It's the language of daily life — markets, homes, music, humour. But it has almost zero representation in AI systems. Most LLMs don't know it exists. Those that do produce gibberish with a French accent.
This isn't just a technical gap. It's an erasure. When AI doesn't speak your language, it doesn't serve your community.
The dataset challenge
The hardest part isn't the training — it's the data. Kreol has:
- No large curated text corpus
- Heavy code-switching with French and English
- Multiple orthographic conventions (the Akademi Kreol Morisien standard is recent)
- Very little digital presence compared to major languages
- Expand the dataset with more diverse sources
- Release a public benchmark for Kreol LLM evaluation
- Open-source the training recipes and evaluation framework
- Deploy a demo for community feedback
We've been collecting from community radio transcripts, public government documents, literature with permission, and social media with consent. We're also working with native speakers to validate and annotate.
Technical approach
We're fine-tuning from a base multilingual model rather than training from scratch. This gives us transfer learning from related languages (French, English) while adapting specifically to Kreol's grammar and vocabulary.
Early benchmarks show promising results on translation and basic comprehension. The model already handles code-switching better than anything publicly available.
What's next
This is a multi-year project. We're not rushing. If you're a Kreol speaker and want to help — especially with data validation — get in touch.