Banglish Embed — MiniLM Small

One space. Two scripts. Half the weight.

কারো কাজ করলে বলবেন প্লিজ
Karo kaj korle bolben please

Search in either. Find both.

Banglish e

ki eta? Banglish diye search korle Bangla lekha o pabe, ulta o hobe. Ei tate shob theke boro lekha dhore — prai shottor shobdo.

ki lagbe? pip install sentence-transformers. Adha GB-r o kom, tai purono computer e o namano jay. Internet lagbe na, account lagbe na.

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

aro choto cai? banglish-embed-tiny — 12 MB.

Use it

from sentence_transformers import SentenceTransformer

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

0.47 GB. Runs on a personal computer's CPU. No GPU, no account, no internet.

Built on paraphrase-multilingual-MiniLM-L12-v2, so it fits pipelines already built around that model.

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. In mistakes: 52 here against 26 for the larger one — half the size costs you about one extra mistake in every eighty searches.

If those mistakes matter to you, use banglish-embed instead. It is the more accurate model. This one is for when a gigabyte is the problem.

If 0.47 GB is still too big

There is a 12 MB one: banglish-embed-tiny. Trained from scratch with its own 16,000-word Banglish vocabulary, which is how it stays that small — and it scores a little higher than this model on the same 2,000 pairs.

The catch: that vocabulary is its own, so it will not slot into a pipeline built around a standard multilingual tokenizer. This one will.

Limits

About seventy words, in either script, then it truncates. What it knows is narrow: whether two sentences mean the same. It will not summarize, answer, or rank anything else.

Fine print

98.8% 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 system, chose the base models, 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. Base model paraphrase-multilingual-MiniLM-L12-v2 (Apache 2.0). Pair data from BanglaTLit (MIT).

@misc{banglishembedminilm2026,
  author = {Sifat Febo},
  title  = {Banglish Embed MiniLM Small: one space, two scripts, half the weight},
  year   = {2026},
  url    = {https://huggingface.co/sifat-febo/banglish-embed-minilm-small}
}