Banglish Embed
One space. Two scripts.
কারো কাজ করলে বলবেন প্লিজ
Karo kaj korle bolben please
Search in either. Find both.
Banglish e
ki eta? Bangla ar Banglish — ek kotha duivabe likhle, model duitake ek jaygay rakhe. Tai Banglish diye search korle Bangla lekha o paba.
ki lagbe? pip install sentence-transformers. 1 GB, CPU tei chole.
Internet lagbe na, account lagbe na, taka lagbe na.
mone rakho: choto lekhar jonno — ek-duita line. Lomba pera'r jonno na.
Use it
from sentence_transformers import SentenceTransformer
model = SentenceTransformer("sifat-febo/banglish-embed")
v = model.encode(["Karo kaj korle bolben please",
"কারো কাজ করলে বলবেন প্লিজ"])
model.similarity(v[0], v[1])
0.95 GB. Runs on a laptop CPU. No GPU. No account. No internet.
Why
We write Banglish. Our data is Bengali. Nothing matches.
Now it does.
How good
One Bengali sentence. 1,999 Banglish ones to choose from.
It picks the right partner first, 199 times out of 200.
Same 2,000 held-out pairs as the previous release. The misses fell from 62 to 26 — more than half of them gone. And we read all 26 by hand: most are test rows that cannot be won (a name anonymized on one side only, two near-identical sentences fighting over first place, one line of corrupted emoji). On clean sentences it is close to unbeaten.
What changed
The previous release was a 12 MB model built from scratch. This one is MuRIL, further pretrained on Banglish and then trained on the same pairs with the same recipe. A gigabyte where the original was 12 MB — and less than half the misses. And now a stock BERT, so any tool that speaks sentence-transformers or plain transformers loads it.
The 12 MB original is still here, pinned to its final commit:
SentenceTransformer("sifat-febo/banglish-embed",
revision="df4fb294a79903d3614c8f813e13d0e8eccd4bdc")
The backbone ships on its own as banglish-encoder — fine-tune it for anything Banglish.
Limits
About fifty words, then it truncates. And it knows exactly one thing: that two sentences mean the same. Nothing more.
Fine print
Both releases were scored on the exact same 2,000 pairs, unwinnable rows included — nothing was dropped to raise a number. Treat 99.5% as what it is: first place against 1,999 distractors, counted the hard way (ties count as misses).
AI Disclosure
Built by Claude Code (Anthropic) under the author's direction and review. Every decision about architecture, data, and what to publish or reject was made by the author.
License
Apache 2.0. Base model MuRIL (Google, Apache 2.0). Pair data from BanglaTLit (MIT).
@misc{banglishembed2026,
author = {Sifat Febo},
title = {Banglish Embed: one space, two scripts},
year = {2026},
url = {https://huggingface.co/sifat-febo/banglish-embed}
}
- Downloads last month
- 56