Deploying LLMs with Hugging Face Jobs

Using Hugging Face jobs and skills to fine-tune LFM with Codex / Claude Code with a SKILL.

This guide covers how to use Unsloth and Liquid LFM2.5 for fast LLM fine-tuning through coding agents like Claude Code. Unsloth provides ~2x faster training and ~60% less VRAM usage compared to standard methods.

You will need

• A Hugging Face account (required for HF Jobs)

• A Hugging Face token with write permissions

• A coding agent (Open Code, Claude Code, Codex)

• Read our Claude Code guide on setting them up.

Installing the Skill

Claude Code

Claude Code discovers skills through its plugin system.

1. Add the marketplace:

/plugin marketplace add huggingface/skills

1. Browse available skills in the Discover tab:

/plugin

1. Install the model trainer skill:

/plugin install hugging-face-model-trainer@huggingface-skills

For more details, see the Claude Code plugins docs and the Skills docs.

Codex

Codex discovers skills through AGENTS.md files and .agents/skills/ directories.

Install individual skills with $skill-installer

$skill-installer install https://github.com/huggingface/skills/tree/main/skills/hugging-face-model-trainer

For more details, see the Codex Skills docs and the AGENTS.md guide.

Quick Start

Once the skill is installed, ask your coding agent to train a model. We're using Liquid LFM2.5

Train LiquidAI/LFM2.5-1.2B-Instruct on trl-lib/Capybara using Unsloth on HF Jobs

The agent will generate a training script based on an example in the skill, submit the training to HF Jobs, and provide a monitoring link via Trackio.

Using Hugging Face Jobs

Training jobs will run on Hugging Face Jobs — fully managed cloud GPUs. If you are familiar with Google Colab credits, Hugging Face Jobs also offers a similar credits system. It is a Pay As You Go structure, or your can get credits beforehand. The agent:

1. Generates a UV script with inline dependencies

2. Submits it to HF Jobs via the hf CLI

3. Reports the job ID and monitoring URL

4. The trained model is pushed to your Hugging Face Hub repository

Example Training Script

The skill generates scripts like this:

# /// script
# dependencies = ["unsloth", "trl>=0.12.0", "datasets", "trackio"]
# ///

from unsloth import FastLanguageModel
from trl import SFTTrainer, SFTConfig
from datasets import load_dataset

model, tokenizer = FastLanguageModel.from_pretrained(
    "Qwen/Qwen2.5-0.5B",
    load_in_4bit=True,
    max_seq_length=2048,
)

model = FastLanguageModel.get_peft_model(
    model,
    r=16,
    lora_alpha=32,
    lora_dropout=0,
    target_modules=["q_proj", "k_proj", "v_proj", "o_proj",
                    "gate_proj", "up_proj", "down_proj"],
)

dataset = load_dataset("trl-lib/Capybara", split="train")

trainer = SFTTrainer(
    model=model,
    tokenizer=tokenizer,
    train_dataset=dataset,
    args=SFTConfig(
        output_dir="./output",
        push_to_hub=True,
        hub_model_id="username/my-model",
        per_device_train_batch_size=4,
        gradient_accumulation_steps=4,
        num_train_epochs=1,
        learning_rate=2e-4,
        report_to="trackio",
    ),
)

trainer.train()
trainer.push_to_hub()

The cost for training with Hugging Face Jobs is below:

Model Size	Recommended GPU	Approx Cost/hr
<1B params	t4-small	~$0.40
1-3B params	t4-medium	~$0.60
3-7B params	a10g-small	~$1.00
7-13B params	a10g-large	~$3.00

For a full overview of Hugging Face space pricing, check out the guide here.

Tips for Working with Coding Agents

• Be specific about the model and dataset to use and include Hub IDs (e.g., Qwen/Qwen2.5-0.5B, trl-lib/Capybara). Agents will search for and validate those combinations.

• Mention Unsloth explicitly if you want it used. Otherwise the agent will decide framework based on model and budget.

• Ask for cost estimates before launching large jobs

• Request Trackio monitoring for real-time loss curves

• Check job status by asking the agent to inspect logs after submission

Resources

• Hugging Face Skills Repository