HuggingFace just made the process of fine-tuning open source LLMs much easier through the use of Skills.

Similarly to Claude's Skills, HuggingFace Skills take care of the boilerplate required to fine tune LLMs. That boilerplate consists of validating your dataset, auto-selecting the best GPU hardware for the job, as well as submitting the job for processing, monitoring the training progress, and finally publishing the tuned model to the HuggingFace Hub.

That process is now condensed into just an instruction of:

Fine-tune Qwen3-0.6B on the open-r1/codeforces-cots dataset for instruction following

For instance, for the 0.6B model used above and on a demo dataset, the agent would select a t4-small—enough GPU as the cheapest option available and choose supervised fine-tuning. The other supported options are Direct Preference Optimization and Reinforcement Learning with verifiable rewards.

After submission, you get back the job's details so that you can monitor its progress, along with the expected time to completion and the estimated cost. Just ask at anytime "How's my training job doing?" to get a fresh view of its progress. Finally after the job finishes it can also be automatically released on the HuggingFace Hub.

This functionality is all packaged as Skills which by definition are Agent Context Protocol (ACP) definitions for AI/ML tasks like dataset creation, model training, and evaluation. HggingFace has borrowed the concepts from Claude, something that we examined recently in "Turn Claude Into Your Personal Research Assistant" :

What are Claude Skills? If you're familiar with CustomGPTs then Skills is something similar. They both enable the LLM to access specialized agents. In the Skills case, they are shaped as plugins which comprise of a mix of instructions, scripts, and resources that Claude can load dynamically when trying to perform a task.

The official HggingFace examples use Claude as the coding Agent, but Codex and Gemini CLI will work just as fine. Integrations for Cursor, Windsurf, and Continue are on the way too.

To get started with the skills and by using Claude Code, you need to:

Register the repository as a plugin marketplace:
/plugin marketplace add huggingface/skills

Install the skill:
/plugin install <skill-folder>@huggingface-skills

Replace <skill-folder> with one of the following:

• hf_dataset_creator/
  Prompts, templates, and scripts for creating structured training datasets
  
  
• hf_model_evaluation/
  Instructions plus utilities for orchestrating evaluation jobs, generating reports, and mapping metrics
  
  
• hf-llm-trainer/
  Comprehensive training skill with SKILL.md guidance, helper scripts (e.g., train_sft_example.py, convert_to_gguf.py, cost estimators)
  
  
• hf-paper-publisher/
  Tools for publishing and managing research papers on Hugging Face Hub. Index papers from arXiv, link papers to models/datasets, generate professional research articles from templates, and manage paper authorship.

All fine, but there's limitations too: the Skills can only be used for models ranging from 1 to 7B of parameters; for large models (7B+), they are not suitable. And, you also need a Pro HuggingFace account.

With that said, if you do meet the requirements, then HF Skills is going to certainly save you time and money. Download and install it from its Githbub repo.

More Information

Hugging Face Skills on Github

We Got Claude to Fine-Tune an Open Source LLM

Related Articles

Turn Claude Into Your Personal Research Assistant

To be informed about new articles on I Programmer, sign up for our weekly newsletter, subscribe to the RSS feed and follow us on Facebook or Linkedin.

Comments

Make a Comment or View Existing Comments Using Disqus

or email your comment to: comments@i-programmer.info