Banglish Embed — Tiny

Twelve megabytes. Both scripts. Runs in a browser tab.

কারো কাজ করলে বলবেন প্লিজ
Karo kaj korle bolben please

Search in either. Find both.

Banglish e

ki eta? Ek kotha Bangla te ba Banglish e — model duitake ek jaygay rakhe. Pura jinish ta 12 MB, tai eta browser tab er bhitorei chole.

ki lagbe? 12 MB. Phone e chole, browser e chole, CPU tei chole. Internet lagbe na, account lagbe na, taka lagbe na.

aro shothik cai? banglish-embed — onek boro, kintu aro bhalo.

Use it

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("sifat-febo/banglish-embed-tiny")
v = model.encode(["Karo kaj korle bolben please",
                  "কারো কাজ করলে বলবেন প্লিজ"])
model.similarity(v[0], v[1])

Or in the browser, with no Python at all:

import { pipeline } from '@huggingface/transformers';
const embed = await pipeline('feature-extraction',
                             'sifat-febo/banglish-embed-tiny',
                             { dtype: 'fp32' });

How good

Given one Bengali sentence and 2,000 Banglish ones, it picks the right partner first about 99 times out of 100.

All three were tested on the same 2,000 pairs.

Eighty times smaller than the big one, and it gives up half a point.

What it is

Trained from scratch — not a fine-tune of anything. Its own 16,000-word vocabulary, built for Bengali script and Banglish together, which is why twelve megabytes is enough: it never had to carry a hundred languages it would not use.

The trade: that vocabulary is its own. It will not drop into a pipeline built around a standard multilingual tokenizer. The other two will.

Runs without Python

onnx/model.onnx is the same weights in a portable format — for browsers, Node, C#, phones. It was checked sentence by sentence against the PyTorch model before publishing: same vectors, to six decimal places.

Limits

About fifty words of Banglish, or about fifteen words of Bengali script, and then it truncates — a 16,000-piece vocabulary spends more of itself on Bengali letters than on Latin ones. And it knows exactly one thing: whether two sentences mean the same. Nothing more.

Fine print

99.0% means the right sentence came back ahead of all 1,999 others, and a tie counts as a mistake, not a win. Nothing was removed from the test set to make that number look better.

AI Disclosure

The author designed the model, chose the data it learned from, ran all training and measurement on the author's own machine, read the models' output, and chose what to publish and what to withhold. Claude Code (Anthropic) was used as a coding and writing tool under that direction.

License

Apache 2.0. Trained from scratch — no base model. Pair data from BanglaTLit (MIT).

@misc{banglishembedtiny2026,
  author = {Sifat Febo},
  title  = {Banglish Embed Tiny: twelve megabytes, both scripts,
            runs in a browser tab},
  year   = {2026},
  url    = {https://huggingface.co/sifat-febo/banglish-embed-tiny}
}