HuggingFace Model Trainer

HuggingFace Model Trainer corresponds to the huggingface-llm-trainer skill in the official huggingface/skills repository (skill folder), renamed from hugging-face-model-trainer in March 2026. It teaches a coding agent to fine-tune language and vision-language models with TRL, or with Unsloth, on Hugging Face Jobs cloud GPUs, so you do not need a local GPU. Results are pushed to the Hub.

It is one of the most detailed skills in the repository, with about ten reference documents and example scripts.

Key Features

• Training methods: SFT, DPO, GRPO and reward modeling, with references/training_methods.md explaining when to use which.
• Unsloth path: the skill recommends Unsloth when GPU memory is tight, models are larger than 13B, or you train vision-language models.
• Example scripts: train_sft_example.py, train_dpo_example.py, train_grpo_example.py and unsloth_sft_example.py, all UV scripts with inline dependencies.
• Helpers: dataset_inspector.py validates a dataset's format before paying for GPUs, estimate_cost.py estimates cost, and convert_to_gguf.py exports for llama.cpp, Ollama or LM Studio.
• Monitoring: every training script should include Trackio.
• Guardrails: push_to_hub=True plus an HF_TOKEN secret, because the job environment is ephemeral; TRL configs use max_length, not max_seq_length.

Use Cases

• Instruction-tuning a small open model on your own chat data.
• Preference tuning with DPO after a first SFT pass.
• Producing a GGUF file to run the result locally.

Pricing

The skill is free (Apache-2.0 repository). Training runs on Hugging Face Jobs, which the skill says requires a Pro, Team or Enterprise plan and bills by GPU type and time.

Getting Started

1. Install HuggingFace CLI, then hf skills add huggingface-llm-trainer.
2. Log in and make sure your token has write access.
3. Validate the dataset first: uv run scripts/dataset_inspector.py --help.
4. Ask for a small demo run (the skill suggests 50 to 100 examples on t4-small) before a real one.

Limitations and risks: the default 30 minute job timeout is too short for most training, and a job that times out loses all progress. The skill tells agents to submit jobs through a Hugging Face MCP hf_jobs() tool; without that server, run the same UV script with hf jobs uv run.

FAQ

Can it train on my own GPU?

The focus is Hugging Face Jobs; a macOS local training reference exists for small experiments.

Does it cover vision detection models?

No. Hugging Face ships a separate huggingface-vision-trainer skill for that.

Alternatives

• Unsloth: the faster fine-tuning library, usable on your own hardware.
• Modal: run your own training code on serverless GPUs.
• HuggingFace Evaluation: benchmark the checkpoint afterward.

Conclusion

The most complete way to let an agent run fine-tuning on Hugging Face infrastructure, with real guardrails around cost and data loss. Pair with experiment tracking. More in the skills hub.