Does this skill send my conversation data to an external server?

No, the entire pipeline runs locally. It processes your Claude history and skill files on your machine and stores them in a local DuckDB instance for training.

How does the skill prepare training data?

It uses DuckDB to parse your Claude JSONL history and local markdown files, automatically generating instruction-output pairs and formatting them into the JSONL chat format required for training.

Which models are recommended for fine-tuning with this skill?

The pipeline is designed for the Qwen2.5 series, with specific recommendations from 0.5B to 32B parameters based on your Mac's available RAM (16GB to 64GB+).

What hardware is required for the local-finetune skill?

This skill is specifically optimized for Apple Silicon (M-series chips) and requires a native arm64 shell environment to run the MLX-based training pipeline.

Local Model Fine-Tuning Pipeline

Name: Local Model Fine-Tuning Pipeline
Author: plurigrid

byplurigrid

•

データサイエンスとML

Automates the creation of high-quality training datasets from Claude conversation history and local skill files to fine-tune LLMs using MLX on Apple Silicon.

概要

The local-finetune skill provides a comprehensive pipeline for embedding domain-specific knowledge and personal interaction styles into local models via LoRA fine-tuning. Optimized specifically for Apple Silicon, it orchestrates a sophisticated data flow using DuckDB to process Claude conversation history and local skill markdown files into structured training pairs. By leveraging the MLX framework, this skill enables developers to create highly specialized, personalized versions of models like Qwen 2.5, bridging the gap between general-purpose AI and custom-tailored coding assistants that understand your specific project conventions.

主な機能

2 GitHub stars
Seamless model fusion and adapter-based inference workflows
Optimized LoRA fine-tuning for Apple Silicon hardware via MLX
Structured data orchestration using DuckDB and ACSets-inspired schemas
Automated data ingestion from Claude history and local skill directories
Built-in utilities for training, validation, and test dataset splitting

ユースケース

Personalizing a local coding assistant to match your project's unique architectural patterns and style guides.
Refining model performance for specific domain-driven design tasks using your own interaction history.
Embedding internal proprietary documentation and niche technical skills into an offline-first LLM.

概要

主な機能

2 GitHub stars
Seamless model fusion and adapter-based inference workflows
Optimized LoRA fine-tuning for Apple Silicon hardware via MLX
Structured data orchestration using DuckDB and ACSets-inspired schemas
Automated data ingestion from Claude history and local skill directories
Built-in utilities for training, validation, and test dataset splitting

ユースケース

Personalizing a local coding assistant to match your project's unique architectural patterns and style guides.
Refining model performance for specific domain-driven design tasks using your own interaction history.
Embedding internal proprietary documentation and niche technical skills into an offline-first LLM.