Modelpublic

tardellirs/embeddinggemma-pt-br-48k

sourceHugging Facegemmaupdated 3mo agoView on Hugging Face

1likes88downloads

Model Card

embeddinggemma-pt-br-48k

A Portuguese-only text-embedding model, vocabulary-trimmed from `google/embeddinggemma-300m` to a 48k token vocabulary (~144M params, MTEB(por) mean_16 0.7098 = 97.8% of the full model at 47% of its size). No training — only the token embedding matrix was sliced; the transformer encoder and pooling/Dense heads are identical to the base model. Produced with 🛠️ [embedding-vocab-trimmer](https://github.com/tardellirs/embedding-vocab-trimmer).

Part of the embeddinggemma-pt-br family — **64k** is the recommended sweet spot:

model	params	`mean_16`	% of full
google/embeddinggemma-300m	~308M	0.7257	100%
embeddinggemma-pt-br-128k	~207M	0.7192	99.1%
embeddinggemma-pt-br	~157M	0.7172	98.8%
embeddinggemma-pt-br-48k ⭐	~144M	0.7098	97.8%
embeddinggemma-pt-br-32k	~131M	0.6881	94.8%
embeddinggemma-pt-br-24k	~125M	0.6895	95.0%
embeddinggemma-pt-br-16k	~119M	0.6520	89.8%

Usage

python

from sentence_transformers import SentenceTransformer
model = SentenceTransformer("tardellirs/embeddinggemma-pt-br-48k")
emb = model.encode(["O Brasil é um país tropical da América do Sul."], normalize_embeddings=True)

Uses EmbeddingGemma's task prompts (prepend task: search result | query: / title: none | text: for retrieval).

Scope

A compression of Google's EmbeddingGemma to Portuguese (deployment/efficiency artifact; data provenance is the base model's). Vocabulary trimming compresses, it does not enhance. Derived under the Gemma license.

Benchmark: MTEB(por) · leaderboard.