Summary

unsloth-buddy automates the end-to-end LLM fine-tuning lifecycle for agents like Claude Code and Gemini CLI. It targets developers needing to fine-tune models for specific tasks (SFT, DPO, GRPO, vision) on diverse hardware, including NVIDIA GPUs and Apple Silicon, offering a zero-friction experience from data prep to deployment.

How It Works

The project orchestrates fine-tuning through an eight-phase process, from environment setup and data formatting to training, evaluation, and export. A key differentiator is its self-evolving memory system, synthesizing lessons from past projects into a local knowledge base (~/.gaslamp/). This memory is injected into new projects, enabling the agent to adapt and improve its recommendations and execution over time for specific user setups.

Quick Start & Requirements

Installation is agent-centric: Claude Code users use /plugin marketplace add TYH-labs/unsloth-buddy, Gemini CLI users gemini extensions install, or clone the repo. It supports NVIDIA GPUs (Unsloth) and Apple Silicon (mlx-tune/trl), with Python 3.12 specified. Free cloud GPU access is via colab-mcp. Docs: gaslamp.dev/unsloth.

Highlighted Details

• Unsloth Performance: Up to 2x faster training and 80% less VRAM usage than standard HuggingFace, producing exact gradients.
• Comprehensive Training Dashboard: Real-time, task-aware visualizations (SFT, DPO, GRPO, Vision) including GPU memory, loss trends, and reward metrics (http://localhost:8080/).
• Automated Demo Builder: Generates static HTML pages comparing base vs. fine-tuned outputs, with customizable themes.
• Integrated Local Deployment: One-command pipeline using llama.cpp for quantized models (GGUF), including benchmarking and launching an OpenAI-compatible server with chat UI.
• Reproducibility Roadbook: Each project generates gaslamp.md detailing decisions, rationales, and ML concepts for end-to-end reproduction.

Maintenance & Community

unsloth-buddy is part of the Gaslamp AI platform and OpenClaw-compatible, allowing seamless integration with agent frameworks. Its self-evolving memory system implies continuous learning and adaptation based on user interactions.

Licensing & Compatibility

Core components, Unsloth and mlx-tune, are MIT licensed, permitting commercial use and integration into closed-source projects.

Limitations & Caveats

The README does not explicitly list limitations. The system's effectiveness relies on past project data for its self-evolving memory to mature. Hardware compatibility (NVIDIA/Apple Silicon) is a primary requirement for optimal performance.