Hugging Face Unveils 'Hugging Face Skills' That Let Claude Train LLMs Directly

Hugging Face has unveiled 'Hugging Face Skills', which helps AI coding agents go beyond simple code writing to autonomously carry out actual LLM fine-tuning workflows.

With the new hf-llm-trainer skill, agents such as Claude Code, OpenAI Codex, and Gemini CLI can handle the following tasks according to user instructions:

• Validating dataset formats and automatically selecting GPU hardware suited to model size
• Submitting training jobs to Hugging Face Jobs, estimating costs, and monitoring progress
• Automatically uploading the resulting model to the Hugging Face Hub after training is complete
• Debugging support for errors that occur during the training process

This feature supports actual production-level training methodologies such as SFT (Supervised Fine-Tuning), DPO (Direct Preference Optimization), and RLHF (Reinforcement Learning from Human Feedback). It can also handle models ranging in scale from 0.5B to 70B parameters, and supports GGUF conversion for local deployment as well as building multi-stage pipelines.

A Hugging Face Pro or Enterprise plan is required to run training jobs.