Custom models for GitHub Copilot are now available in Limited Public Beta for Copilot Enterprise. This new capability lets you fine-tune Copilot to better understand and align with your organization’s unique coding practices, improving the relevance and accuracy of code suggestions across your projects.
What are custom models?
Custom models are large language models (LLMs) that have been fine-tuned using your organization’s codebases. By training a model on your proprietary libraries, specialized languages, and internal coding patterns, Copilot delivers code suggestions that are more context-aware and tailored to your organization’s needs.
During this beta, you can create a custom model using your GitHub repositories. Optionally, you may also enable the collection of code snippets and telemetry from developers’ Copilot prompts and responses to further fine-tune the model. This process closely aligns Copilot’s suggestions with your coding practices, making them more relevant and accurate. As a result, your development teams will spend less time on code reviews, debugging, and manual code adjustments, ultimately boosting team productivity and ensuring more consistent code quality.
Importantly, your data remains entirely yours. It is never used to train another customer’s model, and your custom model is kept private, ensuring full control, security, and privacy.
When to Use Custom Models
Custom models enable you to make Copilot’s suggestions more relevant to your specific needs, which can lead to higher acceptance rates of the code suggested by Copilot among your developers. Consider using custom models in the following scenarios:
-
Enhance Library and API Usage: When your organization relies heavily on custom libraries or APIs that aren’t well-represented in public datasets, a custom model can prioritize these in its suggestions, making it easier for your developers to follow internal standards.
-
Improve Support for Specialized Languages: If your team works with less common or proprietary languages, custom models can make Copilot much more effective. Fine-tuning helps Copilot understand these languages better, reducing friction and improving productivity.
-
Adapt to Evolving Codebases: As your codebase changes, you have full control over when and how often to retrain your custom model. By regularly retraining, you can ensure that Copilot keeps up with the latest coding patterns, so it continues to provide relevant and accurate suggestions.
How to Get Started
-
Sign Up for the Beta:
Sign up here to participate in the Limited Public Beta and make sure your organization is on the Copilot Enterprise plan.
-
Prepare Your Repositories:
Choose the repositories that best reflect your organization’s coding standards. Include those with proprietary libraries, specialized languages, or key internal frameworks to get the most out of fine-tuning. If your enterprise has multiple GitHub organizations, note that only one organization and its repositories can be used for training during this beta.
-
Enable Telemetry Collection:
To further customize your model, consider enabling the collection of code snippets and telemetry related to developers’ prompts and Copilot’s suggestions. This data will be securely collected and used for additional fine-tuning, improving the accuracy and relevance of Copilot’s output for your team. Your data will only be used to enhance your custom model and will not be shared with others. For more details about our data-handling practices, please visit the Trust & Security Center or review GitHub’s data protection agreement.
-
Training and Usage:
After setup, your custom model will be trained using the selected repositories. Once it’s ready, your developers’ IDEs will automatically start using the custom model, which will inform all in-line code completions.
-
Monitoring & Quality Assessment:
Regularly retrain your custom model to keep it aligned with new code and evolving practices. Use the Copilot Usage Metrics API to track metrics like suggestion acceptance rates and see how much it’s improving.
Additional Resources
