Singaraj B
ReportSelf-taught ML engineer based in Tiruppur, Tamil Nadu. Author of morisien-embed, the first embedding model for Kreol Morisien, archived on Zenodo (DOI 10.5281/zenodo.21877805), and of sante-embed, a 1.5B medical retrieval model across nine languages. Merged contributions to MTEB, sentence-transformers, PEFT and Haystack, including MorisienMTBitextMining, the first Mauritian Creole task in MTEB, with model registrations and results published to embeddings-benchmark/results. No degree, no institutional affiliation.
morisien-embed
morisien-embed is a complete, released pipeline for building evaluation coverage for a language that had none: corpus assembly, leakage audit enforced by tests, MTEB task merged upstream, model trained, results published. MIT licensed throughout, trained in 45 minutes on a free Kaggle T4. This grant applies the same pipeline to Indian languages. Measured against the current MTEB task registry: Tamil has 31 tasks and Malayalam 29, but zero reranking and one STS task each. Bhojpuri has 6 and Maithili 7, with zero retrieval, zero STS and zero reranking. Maithili is a scheduled language of India. All-time downloads: 1140 across morisien-embed, morisien-embed-v1.5 and sante-embed. MorisienMTBitextMining is merged into MTEB and runs as part of the multilingual benchmark, with results published to embeddings-benchmark/results. The task-equity and compute-only plans are open-ended sustainability asks for morisien-embed, submitted to FLOSS/fund in November 2026 with no decision yet. The indic-mteb-coverage plan is a separate, scoped request for the Indian-language expansion.
Leak-audited MTEB evaluation coverage for Indian languages
Budget: INR 180000 native-speaker data (Bhojpuri and Maithili, 60 days each at INR 1500/day); INR 60000 Tamil and Malayalam STS annotation; INR 160000 maintainer time (400 hours over 10 months); INR 30000 compute. 56% paid to native speakers. Milestones: 1 Tamil STS and reranking merged into MTEB, leak audit published, model trained; 2 same for Malayalam; 3 Bhojpuri corpus built with paid speakers; 4 Bhojpuri tasks merged, model trained; 5 same for Maithili.
Closing the task equity gap for low-resource languages
Low-resource languages get only 1-3 MTEB task categories against 14 for high-resource ones, and those are mostly bitext mining and classification. Retrieval, semantic similarity, reranking and clustering are empty; no low-resource language appears in MTEB's STS datasets. This builds those missing categories for creoles and other under-served languages, leak-audited and merged upstream into MTEB, with models trained against them. Reproducible on a single free GPU. 12 months. All MIT.
Compute support
GPU compute alone, to train and evaluate models larger than a free Kaggle T4 allows, and to run the full evaluation suite across the new task categories.