GitHub has announced the limited availability of fine-tuned models for Copilot Enterprise customers. With fine-tuned models, customers can adapt the AI experience so that it better suits their specific coding practices and programming languages. GitHub said this fine-tuning improves the relevance and quality of code suggestions.
GitHub called fine-tuned models the "next big leap in customization." By training models on your organization"s codebase and factoring in telemetry about how your programmers use Copilot"s suggestion, the fine-tuned models can deliver better results.
Commenting on the outcomes of using these models, GitHub said:
"Copilot becomes intimately familiar with your modules, functions, rare languages like legacy or proprietary languages, and internal libraries—delivering code suggestions that are not just syntactically correct, but more deeply aligned with your team’s coding style and standards."
To make these models, GitHub uses something called the Low-Rank Approximation (LoRA) method. For organizations, the main benefit of this method is that training is faster and more affordable than traditional fine-tuning techniques. The LoRA method also integrates insights about how your team interacts with suggestions from Copilot.
One of the big concerns about generative AI is data collection. Services like Gemini and ChatGPT record input from general users to improve their services. With GitHub Copilot"s fine-tuned models, the company says data security is included. Your data is always yours and is never used to train another customer"s model, and your custom model is always private, giving you control.
Explaining the training process, GitHub says:
"When you initiate a training process, your repository data and telemetry data are tokenized and temporarily copied to the Azure training pipeline. Some of this data is used for training, while another set is reserved for validation and quality assessment. Once the fine-tuning process is complete, the model undergoes a series of quality evaluations to ensure it outperforms the baseline model. This includes testing against your validation data to confirm that the new model will improve code suggestions specific to your repositories.
If your model passes these quality checks, it is deployed to Azure OpenAI. This setup allows us to host multiple LoRA models at scale while keeping them network isolated from one another. After the process is complete, your temporary training data is removed from all surfaces, and data flow resumes through the normal inference channels. The Copilot proxy services ensure that the correct custom model is used for your developers’ code completions."
These fine-tuned models are now in a limited public beta, and GitHub is gradually onboarding customers from its waitlist. If you would like to join the waitlist, you can do so here, then just wait patiently.
Source: GitHub