mentee-embed: Training Competitive Multilingual Text Embeddings from Scratch for Arabic, English, and Urdu

We present mentee-embed-v1, a 41M-parameter trilingual text embedding model for Arabic, English, and Urdu trained entirely from scratch — no pre trained backbone, no fine-tuning of existing models. Training uses a custom 50K BPE tokenizer and two stages: masked language modeling followed by relational knowledge distillation from multilingual-e5-base. On in-batch retrieval benchmarks, mentee-embed-v1 achieves avg MRR@10 of 0.585, beating all-MiniLM-L6-v2 (0.396) while remaining functional on all three languages. Code, weights, and tokenizer are publicly released.