How to use the LLM Fine-Tuning Expert agent skill for LoRA

What the llm-finetuning skill makes the agent do

An agent skill is a folder containing a SKILL.md file with instructions that the agent reads when a task matches its description. This one covers LoRA, QLoRA, dataset preparation, hyperparameter choices, evaluation and adapter deployment, using the Hugging Face PEFT ecosystem.

With the skill loaded, the agent can help you prepare instruction-response training data, choose settings such as LoRA rank, alpha, target modules, learning rate and number of epochs, plan an evaluation beyond simple loss scores, and save or merge the resulting adapter files. The skill is know-how only: the file itself does not run any training.

Install the llm-finetuning skill

Download the llm-finetuning ZIP from our page and unzip it. The skill ships as a single file, SKILL.md, so the folder you extract should contain that file.

For a personal install, unzip it into ~/.claude/skills/. For a single project, unzip it into the project's .claude/skills/ folder instead.

In the Claude apps you can instead upload the skill under Settings → Capabilities → Skills.

• Personal: ~/.claude/skills/
• Project only: .claude/skills/ inside your project
• Claude apps: Settings → Capabilities → Skills, then upload

How to trigger the skill with natural requests

You do not need a special command. Describe a fine-tuning task in plain language and mention the relevant terms, such as LoRA, QLoRA, adapter, dataset or fine-tuning, so the agent recognises that the skill applies.

Be specific about your situation: the base model you plan to use, the task, the amount and format of your data, and the hardware you have. The more detail you give, the more tailored the guidance will be.

• Plan a LoRA run for a legal-document summarisation model on a single GPU.
• Review my instruction-response dataset for a medical FAQ assistant before I train.
• Which LoRA rank and learning rate should I start with for a 7B model?
• Help me decide whether QLoRA is worth it for my memory limits.
• How do I save a LoRA adapter and merge it into the base model for inference?

Get the most from the skill: dataset, settings and evaluation

The skill stresses that dataset quality matters more than quantity. It recommends formatting data as instruction-response pairs with a consistent template, with a system field for persona or context, an instruction field for the task and a response field for the expected output. It also suggests starting with the strongest base model you can afford to run.

For settings, the skill describes LoRA rank in the range r=8 to 64, alpha typically at 2x rank, and learning rates between 1e-5 and 2e-4 with a cosine schedule. It suggests 1 to 5 epochs, watching for overfitting, and a warmup ratio of 0.03 to 0.1. Treat these as starting points, because the skill warns that different base models respond differently.

For evaluation, ask the agent to combine validation loss, task-specific metrics such as ROUGE for summarisation or exact match for QA, and structured human review on a held-out set. Perplexity alone can miss output quality.

• Keep evaluation examples out of training data (decontaminate).
• Record hyperparameters, dataset versions and base model checkpoints for every run.
• Stop early if outputs become repetitive or templated.

Safety notes: data, scripts and network

The skill declares that it runs no scripts and uses no network access. It lists no required tools or accounts. Actual training, however, needs GPUs and machine learning tooling that you install and run yourself, and the skill does not provide them.

The skill warns against fine-tuning on personally identifiable information, copyrighted content or harmful material without proper review and filtering. Because the agent could otherwise help move such data into a training pipeline, check your data sources and add a review step before any training run, especially for medical, financial or legal records.

• No scripts are run by the skill.
• No network use is declared.
• Review and filter sensitive or copyrighted data before training.

Who should use this skill and who should not

The skill suits machine learning engineers planning or reviewing a LoRA or QLoRA run, teams preparing instruction-response datasets for a domain model, and developers learning the Hugging Face PEFT workflow step by step.

It is not aimed at people who only want to get more out of an existing chatbot. Changing model weights requires GPUs and ML tooling, so if you only need better prompts, a prompting guide will serve you better.