huggingface-llm-trainer

Hugging Face LLM trainer

huggingface-llm-trainer is Hugging Face's agent skill for fine-tuning language and vision models on its cloud GPUs, using TRL or Unsloth on Hugging Face Jobs, with no local GPU needed.

Agent skill by Hugging Face Official · Updated October 2, 2026

Install Hugging Face LLM trainer

Install the skill with the skills CLI

Terminal

npx skills add huggingface/skills --skill huggingface-llm-trainer

Or, with the hf CLI installed, add it the way Hugging Face's README describes

Terminal

hf skills add huggingface-llm-trainer

Or paste this into your coding agent: Install the agent skill huggingface-llm-trainer from github.com/huggingface/skills. A skill can include scripts that run on your computer, so read its source first.

What Hugging Face LLM trainer does

Hugging Face LLM trainer is a skill for training models on Hugging Face Jobs, its managed cloud GPU service. It covers supervised fine-tuning, DPO preference training, GRPO reinforcement learning and reward models with the TRL library, and suggests Unsloth for tight GPU memory, models over 13B parameters or vision-language models. It can also convert a trained model to GGUF for Ollama, LM Studio or llama.cpp.

The agent writes a training script in the uv inline-dependency format, adds Trackio for live monitoring, and submits it as a job right away, then reports the job ID, a monitoring link and an estimated time. Bundled scripts estimate cost, inspect datasets before training and pick base models from benchmarks. A hardware table on 2026-10-01 runs from about 0.75 dollars an hour on a T4 to 10 to 20 dollars an hour for models above 13B.

Two failure modes get loud warnings. The training environment is wiped when a job ends, so the model must be pushed to the Hub with a write token passed as a secret, or all results are lost. And the default 30 minute timeout is too short for most runs, so the agent sets one to two hours or more.

When to use Hugging Face LLM trainer

You want to fine-tune a small open model on your own dataset without owning a GPU.
You want to try DPO or GRPO training and have the job script written for you.
You want a fine-tuned model converted to GGUF to run locally.
You want a cost estimate before starting a training run.

When to pick something else

Free accounts: Hugging Face Jobs require a Pro, Team or Enterprise plan, and GPU time is billed per hour.
Training on your own local GPU: the skill is built around Hugging Face's cloud jobs.

What Hugging Face LLM trainer needs

A Hugging Face account on a Pro, Team or Enterprise plan
A Hugging Face token with write access
The Hugging Face MCP server, or the hf CLI for the hf jobs fallback

Which agents Hugging Face LLM trainer works in

Hugging Face documents Hugging Face LLM trainer for Claude Code, Codex, Gemini CLI, Cursor, Any agent that reads SKILL.md files, through the skills CLI. The open skills CLI also installs it into 78 agents, including Claude Code, Codex, Cursor, Gemini CLI, GitHub Copilot, OpenCode (we listed it with the CLI on October 1, 2026). See where each agent looks for skills.

The skill tells the agent to submit jobs through the hf_jobs tool of the Hugging Face MCP server. Without that server, it falls back to the hf jobs command in the terminal.