unsloth-buddy skill
A skill for Claude Code and Gemini CLI that runs LLM fine-tuning from dataset to GGUF export, on NVIDIA GPUs and Apple Silicon
- 283 ★ on the repository
- Last commit
- MIT
Extended access
Why this level
- The agent runs scripts and training on your machine, creates model folders of tens of gigabytes and installs Python packages into the environment
- The cloud mode sends the dataset and code into a Google Colab session if you switch it on
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add skills/unsloth-buddyDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- The agent runs scripts and training on your machine, creates model folders of tens of gigabytes and installs Python packages into the environment
- The cloud mode sends the dataset and code into a Google Colab session if you switch it on
This is third-party code. Review the repository files before installing.
What it does
unsloth-buddy turns the agent into a model fine-tuning assistant. It asks two questions about the task and the data, then picks a method (SFT, DPO, GRPO, ORPO, KTO, SimPO), a model and a size that fit your hardware. The work then runs in phases: data preparation, environment check, training, a base-versus-fine-tuned comparison, export. On NVIDIA GPUs training goes through Unsloth, on Macs with M1–M4 chips through mlx-tune, for vision models through mlx-vlm, and for GRPO through TRL. For the cloud the skill can drive a free Google Colab GPU through colab-mcp. During training a dashboard opens on localhost:8080 with the loss curve, memory use and time left. Every project gets its own dated folder and a gaslamp.md file that records the decisions made. When a run ends, the skill saves lessons to ~/.gaslamp/: quirks of your hardware, workarounds, ready recipes. The next project picks these notes up automatically. Export targets are GGUF, merged 16-bit weights and Hugging Face Hub, with Ollama and vLLM for serving.
Who it is for. Machine learning engineers and developers with a MacBook or an NVIDIA GPU who need to fine-tune a model on their own data and are unsure which method and size to choose.
Good fit when
- You have 500 support questions and answers in a CSV and need a summarization model trained on a MacBook Air
- Training hits GPU memory limits, and you need to know whether a 7B model with QLoRA fits a 16 GB T4
- The choice between SFT, DPO and GRPO is not obvious and you want an explanation in plain words
- You want to see the fine-tuned model next to the base one, not only the loss number
- A month later someone else or another agent has to repeat the experiment from the gaslamp.md file
Not a fit when
- You want a ready model from Hugging Face without training your own
- Training runs on a multi-node cluster: the skill covers one machine or one cloud GPU
- You need standard Unsloth on Apple Silicon: it does not run there, so the skill uses mlx-tune
- The task comes down to prompts or RAG and the data volume does not justify fine-tuning
Example request
/unsloth-buddy I have 500 customer support Q&As in faq.csv and want a summarization model, but I only have a MacBook AirLimitations
- Verified setups: Apple Silicon with Python 3.12 or older and mlx-tune, Linux and WSL with NVIDIA (CUDA 11.8, 12.1, 12.4 or newer), Windows through conda with Python 3.12.
- The whole repository has to be installed: the skill ships helper scripts and sub-skills, and a lone SKILL.md does not work.
- The skill itself runs train.py and the environment scripts. Training takes hours and tens of gigabytes of disk.
- The Colab cloud mode is switched on only with your consent.
- The package version in SKILL.md is 1.0.3 and the project moves fast; the last commit in the repository is from June 2026.
How to disable. In Claude Code run /plugin uninstall unsloth-buddy and, if you like, remove the marketplace with /plugin marketplace remove. For Gemini CLI use gemini extensions uninstall unsloth-buddy. If you cloned the repository into .agents/skills/unsloth-buddy, delete that folder. Notes from past runs live in ~/.gaslamp/ and can be deleted separately.
FAQ
How do I install unsloth-buddy in Claude Code?
Run /plugin marketplace add TYH-labs/unsloth-buddy, then /plugin install unsloth-buddy@TYH-labs/unsloth-buddy. After that describe what you want to fine-tune and the skill activates by itself. For Gemini CLI there is a gemini extensions install command with the repository URL and the --consent flag.
Can I fine-tune a model on a Mac without an NVIDIA GPU?
Yes. On M1–M4 chips the skill uses mlx-tune for SFT and DPO, mlx-vlm for image training and TRL with PyTorch MPS for GRPO. By the project's table a Mac with 10 GB of memory handles 7B and one with 24 GB handles 13B.
Which training methods does the skill support?
SFT, DPO, GRPO, ORPO, KTO, SimPO and image training for Qwen2.5-VL, Llama 3.2 Vision, Gemma 3 and Gemma 4. The agent picks one after two questions about the task and data and explains the choice in plain language.
What does the skill remember between projects?
After each run it writes three files to ~/.gaslamp/: lessons.md with gotchas it found, skills.md with recipes and user.md with a profile of your hardware. A new project receives this snapshot read-only and applies it without being reminded.
Where does the skill put project files?
In a separate folder with the date in its name, for example customer_faq_sft_2026_03_17. It holds train.py, the data, logs, adapters and gaslamp.md. By the project description the skill does not touch your repository root.
Discussion
No questions or reviews yet.
Related
SkillsThe skill turns your Claude Code sessions into a dataset, trains a LoRA on Tinker and builds a GGUF for local use.
A local MCP server that stores skills as versions and lets improved variants go live only after a human review
Local coding agent with a terminal UI: it plans, drafts several candidates, tests them in a sandbox and repairs, so compact open models can handle code
A collection of practical skills: Claude 5 and GPT-5.6 model routing, English for engineers, Pi Agent setup, YouTube.
A Hugging Face tool: a strong model creates a skill from agent work traces, a weak one uses it and the evaluation shows the gain
A skill guiding an agent to build a small task-specific model on the distil labs platform, from traces to training.
README in short
The README opens with an example request and a session transcript in which one sentence from the user ends with a model in GGUF format. It continues with installation for Claude Code, Gemini CLI and any agent that follows the Agent Skills standard, a table of differences and eight phases with the output files of each. The end covers supported hardware, the training dashboard, local serving through llama.cpp and demo pages with fine-tuning examples.
README badge
Original SKILL.md
--- name: unsloth-buddy description: This skill should be used when users want to fine-tune language models or perform reinforcement learning (SFT, DPO, GRPO, ORPO, KTO, SimPO) using the highly optimized Unsloth library. Covers environment setup, LoRA patching, VRAM optimization, vision/multimodal fine-tuning, TTS, embedding training, and GGUF/vLLM/Ollama deployment. Should be invoked for tasks involving fast, memory-efficient local or cloud GPU training, specifically when the user mentions Unsloth or when hardware limits prevent standard training. license: Complete terms in LICENSE.txt metadata: author: gaslamp version: "1.0.3" category: machine-learning repository: https://github.com/TYH-labs/unsloth-buddy --- # Unsloth Training & Optimization ## Overview You are the `unsloth-buddy`, a specialized AI assistant that helps machine learning practitioners train and optimize large language models (LLMs) using the Unsloth library. **Unsloth provides massive advantages over standard Hugging Face training:** - **Speed**: ~2x faster training speeds. - **Memory**: Up to 80% less VRAM usage (enabling 70B models on a single 80GB GPU, or 8B models on 12GB). - **Exact Math**: 0% loss in accuracy; Unsloth uses exact manual backprop kernels, not approximations. - **Broad Support**: Text, Vision/Multimodal, TTS, Embedding fine-tuning. All RL methods.