matryoshka-embedding-v1

Matryoshka Text Embedding v1

A multilingual text embedding model with Matryoshka Representation Learning, allowing flexible embedding dimensions from 64D to 1024D.

Model Overview

This model implements Matryoshka Representation Learning, enabling you to truncate embeddings to different dimensions while maintaining good performance. This allows you to balance accuracy, speed, and storage based on your specific needs.

Key Features

• Flexible Dimensions: Choose from 7 different embedding sizes (64D, 128D, 256D, 384D, 512D, 768D, 1024D)
• Multilingual Support: Trained on 100+ languages
• Base Architecture: XLM-RoBERTa
• Max Sequence Length: 8192 tokens

Quick Start

Installation

pip install sentence-transformers

Basic Usage

from sentence_transformers import SentenceTransformer

# Load model
model = SentenceTransformer('hasankursun/matryoshka-text-embedding-v1')

# Full precision (1024D)
embeddings = model.encode(["Your text here"])

# Balanced mode (512D) - Recommended for most use cases
embeddings = model.encode(["Your text here"], truncate_dim=512)

# Fast mode (256D) - For high-throughput applications
embeddings = model.encode(["Your text here"], truncate_dim=256)

# Ultra-fast mode (128D) - For real-time applications
embeddings = model.encode(["Your text here"], truncate_dim=128)

Performance Benchmarks

SciFact (Scientific Document Retrieval)

STSBenchmark (English Semantic Similarity)

• Spearman: 0.8506 (1024D)
• Pearson: 0.8381 (1024D)

STS17 (Multilingual Semantic Similarity)

Average Spearman Correlation across languages: 0.8096

Performance by language pair (1024D):

• Spanish (es-es): 0.8808
• English (en-en): 0.8740
• German (en-de): 0.8245
• Korean (ko-ko): 0.8210
• French (fr-en): 0.8157
• Italian (it-en): 0.8152
• Dutch (nl-en): 0.8190
• Arabic (ar-ar): 0.8056
• Turkish (en-tr): 0.7484
• Spanish-English (es-en): 0.7660
• English-Arabic (en-ar): 0.7191

Use Cases

High Accuracy Applications (768D-1024D)

• Scientific literature search
• Legal document retrieval
• Medical information systems

Balanced Production (512D) - Recommended

• General web search
• E-commerce product search
• Content recommendation engines
• Knowledge base retrieval

High-Throughput Systems (256D-384D)

• Real-time search APIs
• Large-scale document indexing
• Social media search

Mobile & Edge Devices (64D-128D)

• Mobile applications
• IoT devices
• Browser-based search
• Resource-constrained environments

Advanced Usage

Semantic Search

import numpy as np
from sentence_transformers import util

# Index documents with 512D (optimal balance)
documents = [
    "Artificial intelligence is transforming healthcare.",
    "Machine learning models require large datasets.",
    "Quantum computing promises exponential speedups."
]

doc_embeddings = model.encode(documents, truncate_dim=512)

# Search with same dimension
query = "How is AI used in medicine?"
query_embedding = model.encode(query, truncate_dim=512)

# Compute similarities
similarities = util.cos_sim(query_embedding, doc_embeddings)
top_result = np.argmax(similarities)

print(f"Most relevant: {documents[top_result]}")

Integration with FAISS

import faiss
import numpy as np

# Create embeddings with 512D
embeddings = model.encode(documents, truncate_dim=512)
embeddings = embeddings.astype('float32')

# Build FAISS index
dimension = 512
index = faiss.IndexFlatIP(dimension)
faiss.normalize_L2(embeddings)
index.add(embeddings)

# Search
query_embedding = model.encode(query, truncate_dim=512).astype('float32')
faiss.normalize_L2(query_embedding.reshape(1, -1))
distances, indices = index.search(query_embedding.reshape(1, -1), k=10)

Technical Details

Architecture

• Base: XLM-RoBERTa transformer encoder
• Embedding Dimensions: 1024 (full) with 7 supported truncation levels
• Max Sequence Length: 8192 tokens
• Vocabulary Size: 250,002 tokens
• Parameters: ~568M

Training

• Technique: Matryoshka Representation Learning
• Languages: 100+ languages
• Max Input Length: 8192 tokens

Model Files

• pytorch_model.bin - Model weights
• config.json - Model configuration
• tokenizer.json - Tokenizer configuration
• matryoshka_config.json - Matryoshka-specific configuration

License

This model is released under the CC-BY-NC-4.0 (Creative Commons Attribution-NonCommercial 4.0 International) license.

See the LICENSE file for full details and acknowledgments.

Acknowledgments

This model builds upon important foundational work:

• XLM-RoBERTa: Base architecture for multilingual representations
• BAAI: For their contributions through RetroMAE and BGE-M3 papers
• Matryoshka Representation Learning: Training methodology (Kusupati et al., 2022)

Citation

If you use this model in your research or application, please cite:

@misc{matryoshka-text-embedding-v1,
  title={Matryoshka Text Embedding v1},
  author={Hasan Kurşun},
  year={2025},
  url={[https://huggingface.co/hasankursun/matryoshka-text-embedding-v1](https://huggingface.co/hasankursun/matryoshka-text-embedding-v1)}
}