minuscule

A Latin-script vocabulary prune of bert-base-multilingual-cased-ner-hrl. Specifically of Xenova/bert-base-multilingual-cased-ner-hrl, the int8 ONNX export of Davlan/bert-base-multilingual-cased-ner-hrl, stripped of the vocabulary a Latin-script deployment can never tokenize.

178.5 MB → 139.6 MB (−21.8%), with identical output for the languages it still covers. No encoder weight was touched and nothing was re-quantized. This is a packaging change, not a distillation — there is no accuracy trade to evaluate.

Why it works

bert-base-multilingual-cased carries a 119,547-token vocabulary spanning 104 languages. At 768 dimensions that embedding table is ~92 MB of the 178 MB file — more than half the model, and every user downloads all of it. If you ship Latin-script languages, the Cyrillic, CJK, Arabic, Hangul, Hebrew, Devanagari, Greek, Armenian and Thai rows are downloaded by everyone and gathered by no one.

Removing them is exact rather than approximate, for two reasons:

• The embedding is per-tensor quantized — a scalar scale and zero-point shared by the whole tensor — and is read by a plain Gather. Selecting a subset of rows preserves both the int8 values and their dequantization. No weight is re-quantized and nothing drifts.
• WordPiece is longest-match over the available pieces. Only pieces unreachable from a kept script were removed, so segmentation of Latin text is unchanged too.

Vocabulary: 119,547 → 68,875 tokens (57.6% kept).

Verification

Not asserted — measured. Both models were run over the same 77 documents — synthetic contracts, financial statements and medical records across 7 languages — comparing raw logits rather than entity lists, because entity lists can agree by luck:

documents	77
entities compared	2,219
segmentation differences	0
entity differences	0
max abs. logit delta	0.0000000000

The comparison is over raw logits and fails on any nonzero difference.

Two limits worth stating plainly. The corpus covers 7 of the 8 kept languages — Latvian is not represented in it, and while the argument above applies to Latvian exactly as it does to the rest, it has not been measured there. And the documents are synthetic: they are shaped like real contracts and records rather than drawn from them.

So: verify on your own corpus before shipping. The guarantee is exact for kept scripts, but "kept" depends on which scripts you selected, and a corpus containing a dropped script will (correctly) diverge.

⚠️ Latin script only

Text in a dropped script is not supported — it degrades to [UNK], not to "slightly worse". Relative to the original model:

• Kept: German, English, Spanish, French, Italian, Latvian, Dutch, Portuguese — and any other Latin-script language the base model handled.
• Removed: Arabic and Chinese, which the original supports and this model does not.

If you need those, use Xenova/bert-base-multilingual-cased-ner-hrl — it is the model this one is built from, unchanged apart from the vocabulary.

The same technique keeps whichever scripts you do ship: retaining kana and ideographs alongside Latin lands at ~151 MB, still 27 MB under the original.

Labels

PER, ORG, LOC, DATE — unchanged from the base model, as B-/I- tags.

Usage (transformers.js)

import { pipeline } from "@huggingface/transformers";

const ner = await pipeline(
  "token-classification",
  "promptshield/minuscule",
  { dtype: "q8" },
);

await ner("Le contrat a été signé par Jean-Baptiste Lefèvre à Bordeaux.", {
  aggregation_strategy: "simple",
});

It also loads in onnxruntime directly — it is a standard BERT token-classification graph with input_ids, attention_mask and token_type_ids.

Provenance and licence

base	google-bert/bert-base-multilingual-cased — Apache-2.0
fine-tune	Davlan/bert-base-multilingual-cased-ner-hrl — AFL-3.0
ONNX export + int8 quantization	Xenova/bert-base-multilingual-cased-ner-hrl — the direct parent of this repo
this repo	vocabulary pruning only — AFL-3.0, inherited

All credit for the model itself belongs to David Adelani (Davlan) for the fine-tune, to the Google Research team for the base model, and to Joshua Lochner (Xenova) for the ONNX export and int8 quantization this repo is built directly on. We changed the vocabulary and nothing else.

The upstream card declares no training datasets; if your use has data-provenance requirements, check with the original authors rather than relying on this card.

Who made this and why

Built for promptShield, an offline document anonymizer. Its web app runs detection entirely in the browser — no document ever leaves the device — so the model is a first-visit download, and 39 MB of unreachable vocabulary was 39 MB charged to every visitor.

Published because the technique generalises: the embedding table is over half of most multilingual encoders, and almost every deployment ships a fraction of the 104 languages.