Morgan-Tanglish-v7 — First Sentence Embedding Model for Romanised Tamil-English (Tanglish)
Tanglish (Romanised Tamil-English code-mixed text) is how ~80 million Tamil speakers actually communicate online — in YouTube comments, WhatsApp, Reddit, Twitter. No usable sentence embedding model existed for it. Until now.
Morgan-Tanglish-v7 is the first purpose-built sentence embedding model for Tanglish.
Benchmark results on TanglishSTS (325 human-annotated pairs, native speaker scored):
ModelSpearmanParamsMorgan-Tanglish-v70.8689118Mall-MiniLM-L6-v20.711022MBAAI/bge-m30.7583570ML3Cube-IndicSBERT0.7642118M
Beats bge-m3 by +11.06 points at 1/5th the parameters. Beats L3Cube-IndicSBERT by +10.47 points.
What makes it different:
The core problem with prior multilingual models on Tanglish is the lexical gap — they treat "nalla iruku", "nalla irukku", and "super ah iruku" as unrelated because they've never seen consistent Tanglish during training. Standard contrastive learning doesn't fix this; it just memorizes surface forms.
Morgan-v7 was trained with Concept Cluster Training — a novel technique where semantically equivalent Tanglish expressions are grouped into concept clusters before pair generation, forcing the model to learn meaning-level representations rather than token-level similarity. Combined with CachedMNRL + CoSENTLoss multi-task training and hard negative mining via Groq/Llama 3.3 70B.
Dataset scale:
185,973 clean Tanglish sentences scraped from YouTube + Reddit (11.8x larger than any prior Tanglish corpus)
86,438+ training pairs with hard negatives
TanglishSTS: first human-annotated STS benchmark for Tanglish
Everything is open-source.
Model: huggingface.co/vishnuexe/Morgan-Tanglish-v7
Corpus: huggingface.co/datasets/vishnuexe/Tanglish-Corpus-185k
Benchmark: huggingface.co/datasets/vishnuexe/TanglishSTS
Demo: huggingface.co/spaces/vishnuexe/Morgan-Tanglish-Demo
Paper: doi.org/10.5281/zenodo.20792177