You can deploy this model and start earning money today!

altaidevorg/
turkish-bge-m3-model2vec

Name: turkish-bge-m3-model2vec
Availability: InStock

sentence-transformers

Views

No views yet

🇹🇷 Turkish BGE-M3 Model2Vec (533k Distilled Champion)

High-performance, ultra-fast Turkish static sentence embedding model distilled from BAAI/bge-m3 on 533k Turkish Wikipedia articles using 256-dimensional PCA teacher representations.

Model Size: 19.36 MB (114x smaller than 2.2 GB BGE-M3)
Embedding Dimension: 256
Vocabulary Size: 39,655 pruned Turkish tokens
Inference Speed: 63,012 sentences/sec on CPU (797x faster than BGE-M3)
Accuracy (STSb-TR gorkem): 91.36% (Preserves ~95% of BGE-M3 teacher accuracy while beating the 182 MB unpruned model's 86.83% baseline)

📊 Benchmark Results

Model	Size	Speed (sent/s)	STSb-TR (`gorkemergune/stsb-tr`) Pearson	STSb-TR (`emrecan/stsb-mt-turkish`) Pearson
👑 BAAI/bge-m3 (Teacher Transformer)	~2,200 MB	79	96.35%	79.57%
🥇 turkish-bge-m3-model2vec (This Model)	19.36 MB	63,012	91.36%	64.34%
🥈 turkish-bge-m3-model2vec-turboquant-4bit	4.92 MB	24,248	91.79%	64.19%
🥉 turkish-bge-m3-model2vec-turboquant-2bit	2.50 MB	20,013	92.19%	63.53%

🚀 Quickstart & Usage

Installation

pip install model2vec

Inference

python

from model2vec import StaticModel

# Load model directly from Hugging Face
model = StaticModel.from_pretrained("altaidevorg/turkish-bge-m3-model2vec")

sentences = [
    "Yapay zeka modelleri doğal dil işlemede çığır açıyor.",
    "Türkiye'nin en kalabalık ve ekonomik merkezi İstanbul'dur.",
    "Hızlı ve hafif vektör modelleri edge cihazlarda çalışabilir."
]

embeddings = model.encode(sentences)
print(f"Shape: {embeddings.shape}")  # (3, 256)

🧠 Distillation Methodology

Vocabulary Pruning: Retains the 39,655 most frequent Turkish subwords from the multilingual BGE-M3 tokenizer.
Teacher Target Reduction: 1024d BGE-M3 teacher embeddings on 533k Turkish Wikipedia articles compressed to 256d via PCA (86.77% variance preserved).
Sentence-Level Weighted Mean Pooling: Eliminates segmentation drift via gradient descent distillation against normalized teacher targets.

📄 License

Apache-2.0