Releases·Tool Release
CodeFinetuner: LoRA fine-tune Qwen2.5-Coder-3B into a local code autocomplete on your own codebase
A developer released CodeFinetuner, an end-to-end open-source pipeline that LoRA fine-tunes small code autocomplete models like Qwen2.5-Coder-3B on a user's own codebase. The resulting GGUF models run fully locally via llama.vim or llama.vscode. It also serves as a reference implementation of a complete LoRA fine-tuning workflow.
49 engagement1 sourceFri, Sep 11, 2026, 06:56 PM
AI Overview
- cuolm's CodeFinetuner pipelines raw code through tree-sitter Structure-Aware FIM parsing, LoRA tuning, and GGUF conversion for llama.vim/llama.vscode
- Supports Mac MPS and NVIDIA CUDA, with optional Unsloth for faster training and lower VRAM usage
- Evaluates via CodeBLEU, edit similarity, exact match, and perplexity; author cautions real usefulness needs editor testing, not just test-set scores
Related
Products·launched
Offline Node.js code review model fine-tuned from Qwen2.5-Coder-1.5B, released as an npm package
A QLoRA fine-tune of Qwen2.5-Coder-1.5B-Instruct that reviews Node.js code diffs fully offline via an npm package, no API key needed. Training data (~740 generated examples) covers real crash stack traces, TypeScript errors, NestJS DI errors, and Prisma codes. It addresses hallucinated bug reports: the base model invented bugs in 9/9 clean diffs while the fine-tune flagged 0/9.
A QLoRA fine-tune of Qwen2.5-Coder-1.5B-Instruct that reviews Node.js code diffs fully offline via an npm package, no API key needed. Training data (~740 generated examples) covers real crash stack traces, TypeScript errors, NestJS DI errors, and Prisma codes. It addresses hallucinated bug reports: the base model invented bugs in 9/9 clean diffs while the fine-tune flagged 0/9.
1 engagement1 sourceredditTue, Oct 6, 2026, 12:08 PM
Social·Community
Local Auto complete code assistant - Vanilla, Fine-tune or RL?
A developer using Qwen2.5-Coder 3B for local inline code completion finds its suggestions too generic and asks the community whether fine-tuning on personal code, reinforcement learning from accepted/rejected suggestions, or per-project fine-tuning works best. The thread highlights growing interest in customizing small local coding models for individual workflows.
A developer using Qwen2.5-Coder 3B for local inline code completion finds its suggestions too generic and asks the community whether fine-tuning on personal code, reinforcement learning from accepted/rejected suggestions, or per-project fine-tuning works best. The thread highlights growing interest in customizing small local coding models for individual workflows.
7 engagement1 sourceredditThu, Sep 3, 2026, 09:00 AM
Social·Community
Users seek local-LLM IDE autocomplete tools
A developer asks for tooling that runs code autocomplete on a local LLM, noting existing options favor cloud models. The question highlights a gap in local-first coding assistance and seeks model recommendations.
A developer asks for tooling that runs code autocomplete on a local LLM, noting existing options favor cloud models. The question highlights a gap in local-first coding assistance and seeks model recommendations.
9 engagement1 sourceredditFri, Aug 14, 2026, 06:25 PM
Social·Community
Guide walks through LoRA fine-tuning of LLMs with LLaMA-Factory 0.9.5/0.9.6
Guide walks through LoRA fine-tuning with LLaMA-Factory 0.9.5/0.9.6
Guide walks through LoRA fine-tuning with LLaMA-Factory 0.9.5/0.9.6
0 engagement1 sourcerssFri, Oct 2, 2026, 10:02 AM
Social·Community
Liquid AI guides GRPO fine-tuning of LFM2.5-350M from 22.6% to 29.7% on IFStruct structured-output benchmark
A practitioner demonstrates fine-tuning a 350M-parameter model for schema compliance using GRPO in just 100 reinforcement learning steps. The post argues structured output is under-measured, since most benchmarks fold schema compliance into broader reasoning or extraction scores rather than scoring it directly. It offers a reusable recipe for teams wiring small LLMs into downstream systems that depend on valid, parseable output.
A practitioner demonstrates fine-tuning a 350M-parameter model for schema compliance using GRPO in just 100 reinforcement learning steps. The post argues structured output is under-measured, since most benchmarks fold schema compliance into broader reasoning or extraction scores rather than scoring it directly. It offers a reusable recipe for teams wiring small LLMs into downstream systems that depend on valid, parseable output.
0 engagement1 sourcerssThu, Sep 3, 2026, 12:00 AM