CodeFinetuner: fine-tune a local code autocomplete model on your own codebase
MountainTop321 · reddit · 2026-09-12
Redditor releases CodeFinetuner, an open-source end-to-end pipeline for LoRA fine-tuning small code autocomplete models (e.g. Qwen2.5-Coder-3B) on your own codebase, outputting GGUF for fully local use via llama.vim/llama.vscode. Supports Mac (MPS) and NVIDIA GPUs, with optional Unsloth.
Pipeline: raw code → tree-sitter parsing into Structure-Aware FIM examples → LoRA fine-tuning → eval (CodeBLEU, edit similarity, exact match, perplexity) → GGUF conversion.
The author cautions that test-set gains don't guarantee editor usefulness (autocomplete tools sample differently from the greedy decoding used in evals), so real-world verification is needed — but it also serves as a complete reference LoRA fine-tuning pipeline.
More from coding & agent
- Dev on Astra: a stubborn genius you shouldn't try to control, and how to use it — kevinkern · 2026-09-12
- Sentry ships Agent Plugin: one command wires AI coding assistants to fix production errors — zeeg · 2026-09-12
- steipete's agent stack: Ghostty over VS Code terminal, GPT-5 for plan reviews — charlieharris01 · 2026-09-12
- Claude Code 2.1.269 ships plugin eval command and prompt tokens jump 69% — ClaudeCodeLog · 2026-09-12
- Claude Code 2.1.269 ships 98 changes: plugin evals and Bash edit diffs — ClaudeCodeLog · 2026-09-12
- Claude Code 2.1.269 released with 98 CLI changes — ClaudeCodeLog · 2026-09-12