Personal Copilot: Train Your Own Coding Assistant

Hugging Face Blog · 2 years ago

Researchers created HugCoder, a code generation model fine-tuned on Hugging Face's public repositories to enable developers to train personalized coding assistants on their own codebases. Using QLoRA parameter-efficient fine-tuning, the 15.5 billion parameter StarCoder model required only 10GB of GPU memory and cost $13.75 to train on a single A100 40GB GPU, compared to $108 for full fine-tuning on eight A100 80GB GPUs. The approach allows organizations to deploy custom code completion models in VS Code through Hugging Face Inference Endpoints while maintaining performance comparable to the base model on standard benchmarks.