rolf-mozilla/multilingual-e5-small-enfr-pruned-q8

multilingual-e5-small — EN+FR vocabulary-pruned, q8 ONNX

Experimental artifact for Mozilla Smart Window embedding evaluation. An int8 ONNX build of intfloat/multilingual-e5-small with the token embedding table reduced to the tokens English and French text actually uses.

Derived from Xenova/multilingual-e5-small's onnx/model_quantized.onnx — the same file Firefox's ML engine pulls today.

What changed

Only the vocabulary. The encoder is byte-for-byte the original.

	original	this build
vocab rows	250,037	18,572
onnx/model_quantized.onnx	118.31 MB	29.43 MB
tokenizer.json	17.08 MB	0.94 MB
total download	135.39 MB	30.37 MB

That is a 4.46x smaller payload. The embedding table was 81.2% of the original weights; the remaining 21.88 MB encoder is untouched, and is a hard floor for this technique.

Why the retained rows are bit-identical

In the source graph the embedding is quantized per tensor (scalar weight_scale / weight_zero_point) and read by a plain Gather on input_ids; the only graph output is last_hidden_state, so there is no LM head and no second vocab-sized tensor. Dropping rows therefore cannot perturb the rows that remain. Token ids were renumbered to the new row indices by filtering the positional Unigram model.vocab list in tokenizer.json; <s>/<pad>/</s>/<unk> keep ids 0-3.

The only behavioural change is re-segmentation: with fewer pieces available, the SentencePiece Unigram model splits some rare words differently.

How the vocabulary was chosen

Token frequencies over ~224k EN+FR texts spanning five domains (MASSIVE intents en/fr, SciFact, SyntecRetrieval, WikiText-2, browser-history retrieval en/fr), keeping every token seen at least 50 times.

The evaluation texts below were held out of that corpus, so the accuracy numbers measure generalization rather than the keep-set fitting its own test.

Accuracy

Per-query retrieval on memory_usage_eval (500 queries, each with 11-21 candidate memories, 2-4 gold; macro-averaged; mean pooling + L2 normalization):

	EN nDCG@5	EN MRR@10	FR nDCG@5	FR MRR@10
original q8	0.7188	0.8665	0.5599	0.7431
this build	0.6963	0.8525	0.5460	0.7298
delta	-0.0225	-0.0140	-0.0139	-0.0133

<unk> rate is 0 on both eval sets and mean sequence length grows ~0.2%, so there is no truncation or latency cost.

A less aggressive keep-set (min count 2, 34,148 rows, 35.0 MB) costs only about -0.003 nDCG@5 but shrinks 3.4x instead of 4x. This repo is the size-favouring end of that trade.

Not for non-Latin input

This build has no rows for CJK, Cyrillic, Arabic, Devanagari or other non-Latin scripts, and the source tokenizer has no byte fallback, so such text degrades to <unk>. It is an EN+FR-only model by construction.

Usage

import { pipeline } from '@huggingface/transformers';
const extractor = await pipeline('feature-extraction', 'rolf-mozilla/multilingual-e5-small-enfr-pruned-q8');
const out = await extractor(['bonjour le monde'], { pooling: 'mean', normalize: true });

Mean pooling, matching the upstream model. No query/passage prefix was used in the evaluation above.

Provenance

Built by scripts/build_pruned_e5_repo.py in Mozilla's multilingual_embeddings_eval bundle. Inherits the MIT license of intfloat/multilingual-e5-small. Temporary evaluation artifact — expect it to be deleted.