Fine-tuning a Coder model with LoRA for specific code generation tasks

This project fine-tunes Qwen2.5-Coder-7B/13B with LoRA adapters to extend its code generation capabilities. Model size is constrained by this notebook to 7B/13B practically because a 35B or larger model would need multi GPU support.

Notes

• The notebook is designed to run in Colab
• It loads training data and stores tuned model on Google Drive
• This script can work with Qwen2.5-Coder-7B/13B but to be able to test in a free T4 instance the base model is replaced to Qwen/Qwen1.5-0.5B *
• A data set is expected in a csv format with columns: 'instruction', 'output' (script will convert it to jsonl format)
• This repository contains a sample dataset (code_dataset.csv)
• The notebook saves the LoRA adapter as well as the merged model to Google Drive, you need enough storage to save the model
• The notebook has a (Test the fine-tuned model) section that loads the merged model from Google Drive

• Qwen/Qwen1.5-0.5 is not a coder model do not expect quality results. It is used to test the script not the code generation quality

Instructions to run fine tuning

• Choose an appropriate GPU runtime: For 0.5B model free T4 is enough. For 7B/13B model an A100 80GB is recommended.
• Upload the sample or your own dataset to Google Drive
• When the script wants to mount your Google Drive provide the necessary permissions
• Run the script
• Once the training is completed and the model is saved to your Google Drive you can run inference only

Instructions to run inference only

• Import the libraries
• Execute the config cell to define the fine tuned model's directory
• Skip the fine-tuning cells and go to "Test the fine-tuned model" section
• Run the cell that defines FineTunedInference class, it also creates an instance
• Run only the last cells in "Test the fine-tuned model" section (modify the prompt variable)