morisien-embed: A Dedicated Text Embedding Model and Benchmark for Mauritian Creole (Kreol Morisien)

Mauritian Creole (Kreol Morisien, ISO 639-3 mfe) is the home language of roughly 90% of Mauritius, yet no dedicated text embedding model exists for it and general multilingual embedders cover it unevenly. We present morisien-embed, to our knowledge the first embedding model built specifically for Mauritian Creole. It is fine-tuned from multilingual-e5-base (278M parameters) on effectively all publicly available Creole–{English, French} parallel text, using hard-negative mining with positive-aware filtering, a cached in-batch contrastive objective, and Matryoshka representation learning. On held-out Mauritian Creole–English bitext retrieval the model reaches 0.9655 nDCG@10 (0.9653 ± 0.0002 over three seeds), against 0.94 for LaBSE, the strongest general model we evaluated, and it leads in all three measured retrieval directions. We also contribute MorisienMTBitextMining, the first Mauritian Creole task in the Massive Text Embedding Benchmark (MTEB), on which the model attains a mean F1 of 0.925 versus 0.848 for LaBSE. We report the model's limitations candidly, including that the demonstrated margin is in-domain, a measured Haitian Creole proximity effect, and a test-visibility note from recipe development. The model, the benchmark dataset and task, and the full training and evaluation code are released openly.