funsloth-train_skill

Overview

This skill generates ready-to-run training notebooks and scripts for fine-tuning models with Unsloth. It helps create SFT, DPO, and GRPO workflows, offering sensible defaults, guided configuration, or interactive ipywidgets notebooks for runtime tuning.

How this skill works

Choose a configuration mode and provide core context (model, dataset, technique, LoRA rank, sequence length, batch size, learning rate, epochs). The generator emits a notebook or script that installs dependencies, verifies the GPU environment, loads the model, applies LoRA, prepares the dataset, runs the selected trainer, and saves the adapter and merged model. It can also produce an interactive notebook with widgets for live configuration.

When to use it

• You want a production-ready fine-tuning notebook with recommended defaults.
• You need step-by-step guidance to pick model, quantization, and training technique.
• You prefer an interactive notebook to configure training at runtime with ipywidgets.
• You need starter scripts for SFT, DPO, or GRPO to run on cloud or local GPUs.
• You want a consistent training structure that includes load, train, save, and quick inference checks.

Best practices

• Start with sensible defaults for quick validation, then iterate hyperparameters based on dev-set results.
• Match model size and quantization to available VRAM using the hardware guide before selecting batch size.
• Use LoRA to reduce checkpoint size and speed up adaptation; experiment with ranks 8–32 for trade-offs.
• Run one epoch first to confirm pipeline and inference, then scale epochs and learning rate conservatively.
• Save both the LoRA adapter and the merged model to support inexpensive inference and full model recovery.

Example use cases

• Generate a production-ready SFT notebook for unsloth/llama-3.1-8b-unsloth-bnb-4bit with defaults to fine-tune on a 100k-example dataset.
• Create a guided notebook that asks about model family, quantization, and training technique to tailor a DPO workflow.
• Produce an ipywidgets notebook for interactive configuration before launching training on RunPod or local GPU.
• Export training scripts for SFT, DPO, or GRPO to run as a Hugging Face Jobs workload.
• Quickly scaffold a training notebook that includes LoRA application, trainer calls, and a small inference test block.