funsloth

内容来源:README.md(说明文档) · 原始地址 · 查看安装指南

原始内容

funsloth: Fully Automated Luxury LLM Fine-Tuning

funsloth logo

A Claude Code skills marketplace for fine-tuning language models with Unsloth. It's premised on the dorkestration paradigm, enabling seamless orchestration of multiple skills to achieve complex workflows using Claude Code.

Overview

funsloth provides six connected skills that guide you through the complete fine-tuning workflow:

funsloth-check → funsloth-train → [hfjobs|runpod|local] → funsloth-upload
Skill Description
funsloth-check Validate datasets, analyze token counts, calculate Chinchilla optimality
funsloth-train Generate Unsloth training notebooks with sensible defaults or custom config
funsloth-hfjobs Train on Hugging Face Jobs cloud GPUs
funsloth-runpod Train on RunPod GPU instances
funsloth-local Train on your local GPU
funsloth-upload Generate model cards and upload to Hugging Face Hub

Installation

claude plugin install funsloth

Or install from source:

git clone https://github.com/chrisvoncsefalvay/funsloth
cd funsloth
claude plugin install .

Usage

Quick start

Just tell Claude what you want to do:

> I want to fine-tune Llama 3.1 8B on my custom dataset

Claude will automatically invoke the appropriate skills.

Manual skill invocation

You can also invoke skills directly:

> /funsloth-check mlabonne/FineTome-100k
> /funsloth-train
> /funsloth-local
> /funsloth-upload

Supported models

Family Sizes Recommended 4-bit
Llama 3.x 1B, 3B, 8B, 70B unsloth/llama-3.1-8b-unsloth-bnb-4bit
Qwen 2.5/3 0.5B-72B unsloth/Qwen2.5-7B-Instruct-bnb-4bit
Gemma 2/3 2B, 9B, 27B unsloth/gemma-2-9b-it-bnb-4bit
Phi-4 14B unsloth/Phi-4-bnb-4bit
Mistral 7B, 8x7B unsloth/mistral-7b-instruct-v0.3-bnb-4bit
DeepSeek 7B+ unsloth/DeepSeek-R1-Distill-Qwen-7B-bnb-4bit

Supported training techniques

  • SFT - Supervised Fine-Tuning
  • DPO - Direct Preference Optimization (Rafailov et al., NeurIPS 2023)
  • GRPO - Group Relative Policy Optimization (DeepSeekMath, 2024)
  • ORPO - Odds Ratio Preference Optimization (Hong et al., EMNLP 2024)
  • KTO - Kahneman-Tversky Optimization (Ethayarajh et al., ICML 2024)

Data formats

Format Structure Use Case
Raw Corpus {"text": "..."} Continued pretraining
Alpaca {"instruction", "input", "output"} Instruction tuning
ShareGPT [{"from": "human", "value": "..."}] Conversations
ChatML [{"role": "user", "content": "..."}] Native chat

Unsloth supports many chat templates including llama-3, chatml, mistral, gemma, phi-3, phi-4, qwen-2.5, alpaca, zephyr and vicuna. Custom templates can be provided as a (template, eos_token) tuple.

Requirements

  • Claude Code CLI
  • Python 3.10+
  • CUDA-capable GPU (for local training)
  • Hugging Face account (for dataset/model hosting)

License

MIT

Acknowledgments