CodeFinetuner is a new pipeline designed to fine-tune local code completion models on specific codebases. Developed over several months, it supports models like Qwen2.5-Coder-3B and can run on Mac (MPS) or NVIDIA GPUs (CUDA), with optional Unsloth integration for enhanced training speed and reduced VRAM usage. The pipeline processes raw code into structure-aware examples, performs LoRA fine-tuning, evaluates the results using metrics like CodeBLEU, and converts the model to GGUF format for local inference. AI
IMPACT Enables developers to create highly customized, locally run code completion models tailored to their specific projects.
RANK_REASON The cluster describes a new open-source tool for fine-tuning local code models.
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →