Qodo-Embed-M-1-1.5B-M2V-Distilled

This project optimizes the Qodo-Embed-1-1.5B model using Model2Vec, reducing its size and dramatically improving inference speed while maintaining most of its performance capabilities.

Overview

Qodo-Embed-1-1.5B is a state-of-the-art code embedding model designed for retrieval tasks in the software development domain. While powerful, it can be resource-intensive for production use cases.

Model2Vec is a technique to distill large sentence transformer models into small, fast static embedding models. This project applies Model2Vec to create an optimized version of Qodo-Embed-1-1.5B with the following benefits:

• Smaller Size: Reduces model size by a factor of 25x
• Faster Inference: Up to 112x faster inference
• Low Resource Requirements: Minimal memory footprint and dependencies
• Maintains Performance: Retains most of the original model's capabilities

Model Information

• Model Name: Qodo-Embed-M-1-1.5B-M2V-Distilled
• Original Model: Qodo-Embed-1-1.5B
• Distillation Method: Model2Vec
• Original Dimensions: 1536
• Distilled Dimensions: 384
• Explained Variance: ~85%
• Size Reduction: 25.26x (from 5.9GB to 233MB)
• Speed Improvement: 112.14x faster

Installation

First, ensure you have the required dependencies:

# Install the base package
uv sync

Usage

Distillation

To create a distilled version of Qodo-Embed-1-1.5B:

uv run python distill.py --pca_dims 384

Options:

• --model_name - Source model name (default: "Qodo/Qodo-Embed-1-1.5B")
• --output_dir - Where to save the distilled model (default: "models/qodo_embed_m2v")
• --pca_dims - Dimensions for PCA reduction; smaller values create faster but less accurate models (default: 384)
• --save_to_hub - Push the model to HuggingFace Hub
• --hub_model_id - Model ID for HuggingFace Hub (required if saving to hub)
• --skip_readme - Skip generating README file (default: True)

Evaluation

To evaluate the distilled model against the original:

uv run python evaluate.py

Options:

• --original_model - Original model name (default: "Qodo/Qodo-Embed-1-1.5B")
• --distilled_model - Path to the distilled model (default: ".")
• --output_dir - Where to save evaluation results (default: "evaluation")

Example Code

from model2vec import StaticModel
from sentence_transformers import SentenceTransformer
import time

# Sample code for embedding
code_samples = [
    "def process_data_stream(source_iterator):",
    "implement binary search tree",
    "how to handle memory efficient data streaming",
    """class LazyLoader:
        def __init__(self, source):
            self.generator = iter(source)
            self._cache = []"""
]

# Load original model
print("Loading original model...")
original_model = SentenceTransformer("Qodo/Qodo-Embed-1-1.5B")

# Load distilled model
print("Loading distilled model...")
distilled_model = StaticModel.from_pretrained("models/qodo_embed_m2v")

# Compare embedding speed
print("\nGenerating embeddings with original model...")
start = time.time()
original_embeddings = original_model.encode(code_samples)
original_time = time.time() - start
print(f"Original model took: {original_time:.4f} seconds")

print("\nGenerating embeddings with distilled model...")
start = time.time()
distilled_embeddings = distilled_model.encode(code_samples)
distilled_time = time.time() - start
print(f"Distilled model took: {distilled_time:.4f} seconds")
print(f"Speed improvement: {original_time/distilled_time:.2f}x faster")

print(f"\nOriginal embedding dimensions: {original_embeddings.shape}")
print(f"Distilled embedding dimensions: {distilled_embeddings.shape}")

Results

The distilled model achieves:

• 25.26x reduction in model size (from 5.9GB to 233MB)
• 112.14x increase in inference speed
• 85.1% explained variance with PCA reduction to 384 dimensions

Detailed evaluation results, including similarity plots and performance metrics, are saved to the evaluation output directory.

Project Structure

• distill.py - Script to create the distilled model
• evaluate.py - Script to compare performance with the original model
• example.py - Example usage of the distilled model
• evaluation/ - Directory containing evaluation results and visualizations

Acknowledgments

This project is built upon the following technologies:

• Qodo-Embed-1-1.5B - The original code embedding model developed by QodoAI
• Model2Vec - The distillation technique used to optimize the model

License

This model is licensed under the QodoAI-Open-RAIL-M license, the same as the original Qodo-Embed-1-1.5B model. Any derivative model must include "Qodo" at the beginning of its name per the license requirements.