gomyk/gemma-student-gemma_emb_compressed

gemma_emb_compressed

Compact multilingual sentence encoder compressed from google/embeddinggemma-300m (24x compression).

Model Details

Property	Value
Base model	google/embeddinggemma-300m
Architecture	gemma3_text (decoder)
Hidden dim	384 (from 768)
Layers	4 (from 24)
Intermediate	576
Attention heads	1
KV heads	1
Vocab size	19,485 (from 262,144)
Parameters	~12.5M
Model size (FP32)	47.7MB
Compression	24x
Distilled	No

Quick Start

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("gemma_emb_compressed", trust_remote_code=True)

sentences = [
    "Hello, how are you?",
    "안녕하세요, 잘 지내세요?",
    "こんにちは、元気ですか？",
    "你好，你好吗？",
]

embeddings = model.encode(sentences)
print(embeddings.shape)  # (4, 384)

MTEB Evaluation Results

Overall Average: 27.12%

Task Group	Average
Classification	37.65%
Clustering	27.39%
STS	17.52%

Classification

Task	Average	Details
AmazonCounterfactualClassification	59.01%	en: 62.78%, en-ext: 61.44%, de: 57.88%, ja: 53.94%
Banking77Classification	19.07%	default: 19.07%
ImdbClassification	52.55%	default: 52.55%
MTOPDomainClassification	38.89%	es: 44.84%, th: 43.67%, de: 41.01%, en: 39.61%, fr: 39.45%
MassiveIntentClassification	22.16%	zh-CN: 31.72%, th: 28.37%, zh-TW: 28.22%, vi: 26.23%, sv: 26.17%
MassiveScenarioClassification	23.12%	zh-CN: 33.22%, zh-TW: 28.98%, ko: 28.13%, th: 27.83%, km: 26.79%
ToxicConversationsClassification	50.12%	default: 50.12%
TweetSentimentExtractionClassification	36.28%	default: 36.28%

Clustering

Task	Average	Details
ArXivHierarchicalClusteringP2P	45.54%	default: 45.54%
ArXivHierarchicalClusteringS2S	45.1%	default: 45.1%
BiorxivClusteringP2P.v2	8.07%	default: 8.07%
MedrxivClusteringP2P.v2	19.06%	default: 19.06%
MedrxivClusteringS2S.v2	17.57%	default: 17.57%
StackExchangeClustering.v2	41.53%	default: 41.53%
StackExchangeClusteringP2P.v2	33.43%	default: 33.43%
TwentyNewsgroupsClustering.v2	8.85%	default: 8.85%

STS

Task	Average	Details
BIOSSES	-0.64%	default: -0.64%
SICK-R	30.8%	default: 30.8%
STS12	23.59%	default: 23.59%
STS13	19.19%	default: 19.19%
STS14	11.24%	default: 11.24%
STS15	30.55%	default: 30.55%
STS17	12.2%	es-es: 44.48%, en-en: 34.75%, ko-ko: 33.59%, ar-ar: 20.35%, en-ar: 10.34%
STS22.v2	15.19%	zh: 44.99%, es: 35.77%, es-en: 29.78%, pl-en: 28.62%, ar: 26.32%
STSBenchmark	15.56%	default: 15.56%

Training

Created via multi-method model compression (no additional training):

1. Teacher: google/embeddinggemma-300m (24L, 768d, 303M params)
2. Layer pruning: 24 → 4 layers (uniform selection)
3. Hidden dim: 768 → 384
4. Vocab pruning: 262,144 → 19,485 (90% cumulative frequency)
5. Compression ratio: 24x

License

This model is a derivative of Google's Gemma. Gemma is provided under and subject to the Gemma Terms of Use found at ai.google.dev/gemma/terms. Use of this model must comply with the Gemma Prohibited Use Policy.

Supported Languages (18)

ko, en, ja, zh, es, fr, de, pt, it, ru, ar, hi, th, vi, id, tr, nl, pl