LK-maker-007/morisien-embed

The first dedicated embedding model for Mauritian Creole (Kreol Morisien): training pipeline, leak-audited benchmark, evaluation # morisien-embed To our knowledge, the first dedicated embedding model for **Mauritian Creole (Kreol Morisien)** — the home language of roughly 90% of Mauritius (2022 census), which general multilingual embedding models don't reliably cover. **Model:** Singaraj/morisien-embed · fine-tuned from multilingual-e5-base on effectively all publicly available Creole parallel data. **Paper:** morisien-embed on Zenodo (DOI 10.5281/zenodo.21877805). ## Usage ```python from sentence_transformers import SentenceTransformer model = SentenceTransformer("Singaraj/morisien-embed") creole = ["Mo pe al bazar aster.", "Bann zanfan pe zwe dan lakour."] english = ["I am going to the market now.", "The children are playing in the yard."] similarity = model.similarity(model.encode(creole), model.encode(english)) ``` No prompt or prefix is required. Trained with Matryoshka loss, so embeddings can be truncated for faster search at a small accuracy cost: `SentenceTransformer("Singaraj/morisien-embed", truncate_dim=256)`. ## Results — Creole→English retrieval, held-out MorisienMT test (1,000 queries) | model | params | ndcg@10 | acc@1 | |---|---|---|---| | paraphrase-multilingual-MiniLM-L12-v2 | 118M | 0.16 | 0.10 | | BAAI/bge-m3 | 568M | 0.46 | 0.36 | | intfloat/multilingual-e5-small | 118M | 0.54 | 0.42 | | intfloat/multilingual-e5-base | 278M | 0.64 | 0.53 | | intfloat/multilingual-e5-large | 560M | 0.73 | 0.65 | | sentence-transformers/LaBSE | 470M | 0.94 | 0.91 | | **morisien-embed** | **278M** | **0.9655** | **0.9440** | - Stable across 3 seeds: ndcg@10 **0.9653 ± 0.0002**. - Creole→French: **0.9751** vs LaBSE's 0.9475. - English→Creole (reversed direction): **0.9588** vs LaBSE's 0.9247. - FLORES+ `mfe` (independent domain, 1,012 unseen sentences): perfect 1.0000 retrieval — though LaBSE also sits at that ceiling (0.9996), so the out-of-domain comparison is saturated rather than won. - E5 baselines were ablated with and without their `query:`/`passage:` prompts on an earlie …