funsloth-train
Generate Unsloth training notebooks and scripts. Use when the user wants to create a training notebook, configure fine-tuning parameters, or set up SFT/DPO/GRPO training.
When & Why to Use This Skill
This Claude skill automates the creation of high-performance LLM fine-tuning environments using the Unsloth library. It generates optimized Jupyter notebooks and Python scripts for various training methodologies including SFT, DPO, and GRPO. By providing hardware-aware configurations and sensible defaults, it significantly reduces the complexity of setting up efficient, memory-optimized training pipelines for models like Llama 3.1, Mistral, and DeepSeek.
Use Cases
- Rapidly generating production-ready SFT (Supervised Fine-Tuning) notebooks with optimized 4-bit quantization for Llama 3.1 or Qwen models.
- Setting up Direct Preference Optimization (DPO) or GRPO training scripts to align models with human preferences or reasoning tasks.
- Configuring memory-efficient training parameters (LoRA rank, alpha, and quantization) specifically tailored to available GPU VRAM (e.g., 8GB, 16GB, or 24GB).
- Creating interactive training templates with ipywidgets for researchers who need to adjust hyperparameters dynamically during runtime.
- Streamlining the transition from local experimentation to cloud-based training on platforms like Hugging Face Jobs or RunPod.
| name | funsloth-train |
|---|---|
| description | Generate Unsloth training notebooks and scripts. Use when the user wants to create a training notebook, configure fine-tuning parameters, or set up SFT/DPO/GRPO training. |
Unsloth Training Notebook Generator
Generate training notebooks for fine-tuning with Unsloth.
Quick Start
Copy and customize the template notebook:
notebooks/sft_template.ipynb
Or use a training script directly:
python scripts/train_sft.py # Supervised fine-tuning
python scripts/train_dpo.py # Direct preference optimization
python scripts/train_grpo.py # Group relative policy optimization
Configuration Modes
Ask the user which mode they prefer:
- Sensible defaults - Production-ready notebook with recommended settings
- Guide me - Walk through each option with explanations
- Leave it empty - Notebook with ipywidgets for runtime configuration
Mode 1: Sensible Defaults
Use these production-ready defaults:
| Parameter | Default | Reasoning |
|---|---|---|
| Model | unsloth/llama-3.1-8b-unsloth-bnb-4bit |
Good balance |
| Max seq length | 2048 | Covers most use cases |
| Load in 4-bit | True | 70% VRAM reduction |
| LoRA rank | 16 | Good trade-off |
| Batch size | 2 | Works on 8GB+ VRAM |
| Gradient accumulation | 4 | Effective batch of 8 |
| Learning rate | 2e-4 | Unsloth recommended |
| Epochs | 1 | Often sufficient |
Mode 2: Guide Me
Ask questions in order. See MODEL_SELECTION.md for model options and TRAINING_METHODS.md for technique details.
Key Questions
- Model family: Llama, Qwen, Gemma, Phi, Mistral, DeepSeek?
- Model size: Based on VRAM (see HARDWARE_GUIDE.md)
- Training technique: SFT, DPO, GRPO, ORPO, KTO?
- Quantization: 4-bit (recommended), 8-bit, 16-bit?
- LoRA rank: 8, 16, 32, 64?
- Sequence length: 512, 1024, 2048, 4096?
- Batch size: 1, 2, 4, 8?
- Learning rate: 1e-5, 5e-5, 2e-4, 5e-4?
- Training duration: 1 epoch, 3 epochs, or specific steps?
Mode 3: ipywidgets
Generate a notebook with interactive configuration widgets. Users select options at runtime.
Notebook Structure
Generate notebooks with these sections:
- Title and Overview - What the notebook does
- Installation - Install Unsloth
- Imports and GPU Check - Verify environment
- Configuration - All tunable parameters
- Load Model - FastLanguageModel.from_pretrained()
- Apply LoRA - FastLanguageModel.get_peft_model()
- Load Dataset - Format-appropriate loading
- Training - SFTTrainer/DPOTrainer/GRPOTrainer
- Save Model - LoRA adapter + merged model
- Test Inference - Quick verification
After Generation
Ask where to run training:
- Hugging Face Jobs - Cloud GPUs (
funsloth-hfjobs) - RunPod - Flexible GPU rentals (
funsloth-runpod) - Local - Your own GPU (
funsloth-local)
Context to Pass
notebook_path: "./training_notebook.ipynb"
model_name: "unsloth/llama-3.1-8b-unsloth-bnb-4bit"
dataset_name: "mlabonne/FineTome-100k"
technique: "SFT"
lora_rank: 16
max_seq_length: 2048
batch_size: 2
learning_rate: 2e-4
num_epochs: 1
Bundled Resources
- notebooks/sft_template.ipynb - Ready-to-use SFT template
- scripts/train_sft.py - SFT script template
- scripts/train_dpo.py - DPO script template
- scripts/train_grpo.py - GRPO script template
- references/MODEL_SELECTION.md - Model recommendations
- references/HARDWARE_GUIDE.md - VRAM requirements
- references/TRAINING_METHODS.md - SFT vs DPO vs GRPO