morisien-embed, Mauritian Creole embeddings
A text embedding model for Kreol Morisien. It measures the meaning of a sentence, so sentences that mean the same thing sit close together, even in different languages. Type below and see for yourself.
Enter any two sentences, Creole, English, French, or a mix, and see how close their meanings are.
Type a Creole sentence; the model finds the closest match among the candidates. Edit the candidates to try your own.
Good to know
What this is (and is not)
It reads a sentence and produces a vector, numbers that stand for the sentence's meaning. Two sentences with the same meaning get near-identical vectors, which is what powers semantic search and cross-language matching. It is not a chatbot and not a translator: it never writes text back. It is the understanding layer that sits underneath such tools.
Limitations, please read before judging a result
- Not native-perfect. On the formal-register benchmark it ranks the correct match first about 94% of the time, which is strong but below a native speaker.
- Register. The training data comes from books translated from English, the Bible among them, plus basic sentences written by hand. Street Creole, slang and SMS-style spelling are untested.
- This demo runs a compressed copy of the model so it can load in your browser. The embedding table is 8-bit and the rest is 16-bit, which costs about two queries in a thousand: accuracy@1 is 0.9440 here against 0.9460 for the full weights. For exact numbers, use the model from the repository.
- This is morisien-embed-v1.5, fine-tuned from LaBSE on 35,064 Creole pairs from MorisienMT and Kreyol-MT. That is not all the public Creole parallel text: google/smol holds 2,472 further pairs the training set does not contain.
Feedback, I would genuinely like to hear from you
This model is for the Mauritian Creole community, and your feedback is what makes it better, especially from native speakers. Tried it and it got something wrong, a bad match, an odd score, slang it misunderstands? Please tell me; that is exactly the signal the next version needs. Questions and doubts are welcome too. Leave a note in the Community tab of this Space, or open an issue on GitHub. No feedback is too small.