1. Habit
  2. Skills
  3. Hugging Face LLM Trainer skill
huggingface · 2K installs

Hugging Face LLM Trainer skill
Runs by itself, training run requested.

Fine-tuned models saved to Hub.

Install the Hugging Face LLM Trainer skill and make it a habit: follow https://gethabit.dev/skills/huggingface/huggingface-llm-trainer.md

  1. 1Copy
  2. 2Paste into Claude Code, Codex or Cursor
  3. 3It runs by itself

2K installs11K starsApache-2.0 licenceSource on GitHub ↗

Hugging Face LLM Trainer

by huggingface · 2K installs

The skill

The hugging face llm trainer skill fine-tunes language and vision models with TRL or Unsloth on Hugging Face Jobs, with SFT, DPO, GRPO, dataset validation, Trackio monitoring and GGUF conversion, for builders without a local GPU.

Only when you ask.

The habit

Fine-tuned models saved to Hub.

↻ Runs by itselftraining run requested

How the habit works.
You don't have to.

Three questionsWhen does it start? How does it check its work? What does it remember?

01

It starts by itself

training run requested. Runs in Claude Code.

  • ↻ training run requested
  • on request to train or fine-tune a model

02

It checks its work

dataset ready, push set. If a check fails, it stops.

  • ✓dataset inspector shows READY or mapping applied
  • ✓push_to_hub=True and hub_model_id are set
  • ✓secrets include HF_TOKEN
  • ✓timeout is set above the default

03

It remembers

training job log. The next run reads it first: past job IDs, hardware, timeouts and failure modes.

loops/hf-finetune-run-loop/log.md

One run, step by step.
Then it stops.

Stop ruleStops after one submitted job per run; no automatic status polling.

  1. 01

    Check the login

    Confirm authentication with hf_whoami and a token with write access.

  2. 02

    Validate the dataset

    Run the dataset inspector and apply mapping code if it says NEEDS MAPPING.

  3. 03

    Pick model and hardware

    Use hf_benchmarks.py for a base model and match the flavor to model size.

  4. 04

    Offer an estimate

    Offer estimate_cost.py for time and a recommended timeout with buffer.

  5. 05

    Write the script

    Base it on train_sft_example.py or a DPO or GRPO template, with Trackio.

  6. 06

    Submit the job

    Submit inline with hf_jobs, with timeout and HF_TOKEN in secrets.

  7. 07

    Report and log

    Give the job ID and monitoring URL, then wait for the user to ask for status.

What it may do.
Nothing more.

Safe by defaultIt writes only in its own folder and asks you before anything goes out.

Is it for you?
Use it when it fits.

Quick checkRead both lists. If the left one sounds like you, the Hugging Face LLM Trainer skill helps.

Use it when

  • ✓ You fine-tune without a local GPU
  • ✓ You run SFT, DPO or GRPO
  • ✓ You need a GGUF export

Skip it when

  • – You train on your own GPUs
  • – Your account has no Jobs access

Questions

Hugging Face LLM Trainer skill, answered.

Why must it push to the Hub?

The Jobs environment is temporary, so an unpushed model is lost when the job ends.

Can it make a GGUF file?

Yes. After training it can convert the model to GGUF with the conversion guide.

How do I install the Hugging Face LLM Trainer skill?

Copy the prompt at the top into your agent. It installs the skill from huggingface/skills and sets up the habit. Or run npx skills add huggingface/skills --skill huggingface-llm-trainer.

Is the Hugging Face LLM Trainer skill free?

Yes. It is published under Apache-2.0. Habit is free too.

Hugging Face LLM Trainer. Now as a habit.

Install the Hugging Face LLM Trainer skill and make it a habit: follow https://gethabit.dev/skills/huggingface/huggingface-llm-trainer.md

  • Free
  • No account
  • Undo in one line

All huggingface skills →More ops skills →

Works in

  • Claude Code
  • Codex
  • Cursor
  • ChatGPT
  • Claude app

Made by huggingface, published under Apache-2.0. Habit links to the original and never changes it. Is this your skill?

Copying didn't work here. Select the line and copy it by hand.