Caslow project

Edge-Optimized Model Fine-Tuning for Code Translation

In this project, I fine-tuned a compact language model (Llama-3.2-3B) to translate Fortran code into Rust, enabling efficient deployment on edge devices. By leveraging GPT-4 insights and knowledge distillation, I reduced trainable parameters by 99.8% using LoRA adapters. The optimized model was deployed on Hugging Face, complete with a Gradio-powered web interface and API for seamless user interaction. This work demonstrates the potential of small language models for high-accuracy code translation in resource-constrained environments.
Github Repo Link: github.com/CodeTranslatorLLM/LinguistLLM/tree/main

Project Results

Reduced trainable parameters by 99.8% using LoRA-based fine-tuning of Llama-3.2-3B, enabling edge-device deployment for Fortran-to-Rust translation.
Deployed the optimized model on Hugging Face with a Gradio-powered web interface and API for seamless user interaction.
Demonstrated the potential of small language models to perform complex code translation tasks efficiently in resource-constrained environments.

Model Fine-Tuning

I gained expertise in optimizing large language models by reducing their complexity and training them for specific tasks, improving efficiency and performance on edge devices.

Knowledge Distillation

This project taught me how to transfer knowledge from a large, powerful model (GPT-4) to a smaller, more efficient model (Llama-3.2-3B), maintaining accuracy while reducing computational requirements.

Deployment and API Integration

I learned how to deploy machine learning models on platforms like Hugging Face and integrate them with user-friendly interfaces and APIs, making them accessible for real-world applications.