Model Overview Card: text-embedding-3-small

Basic Information

• Model Name: text-embedding-3-small
• Developer/Creator: OpenAI
• Release Date: January 25, 2024
• Version: text-embedding-3-small
• Model Type: Text Embedding

Description

• Overview:text-embedding-3-small is an efficient and compact embedding model designed to enhance performance over its predecessor, text-embedding-ada-002. It transforms text into numerical representations that can be easily processed by machine learning models.
• Key Features:
  • Improved Performance: Achieves higher scores on benchmarks for multi-language retrieval (MIRACL) and English tasks (MTEB).
  • Cost Efficiency: Offers a 5x reduction in cost compared to text-embedding-ada-002.
  • Compact Size: Embedding size of 512 dimensions, suitable for memory and storage-constrained environments.
• Intended Use:
  • Search: Enhance search algorithms by ranking results based on relevance.
  • Clustering: Group similar text strings for data analysis.
  • Recommendations: Suggest related items based on text similarity.
  • Anomaly Detection: Identify outliers in data.
  • Diversity Measurement: Analyze the diversity of text data.
  • Classification: Classify text strings by their most similar labels.
• Language Support:Supports multiple languages, improving accessibility and usability across diverse linguistic datasets.

Technical Details

• Architecture:Utilizes a transformer-based architecture optimized for efficiency and performance.
• Training Data:Trained on a diverse set of text sources to capture a wide range of linguistic patterns and semantics.
• Data Source and Size:Extensive dataset comprising millions of text documents, ensuring a broad understanding of language.
• Diversity and Bias:Training data is selected to minimize bias and ensure robust performance across different demographics and use cases.

Performance Metrics

• Comparison to Other Models:
  • MIRACL Score: Improved from 31.4% (ada-002) to 44.0%.
  • MTEB Score: Increased from 61.0% (ada-002) to 62.3%.
• Accuracy:Demonstrates higher accuracy in both multi-language and English-specific benchmarks.
• Speed:More efficient than previous models, reducing latency and computational requirements.
• Robustness:Handles diverse inputs effectively, ensuring reliable performance across various applications.