用 Unsloth 与 Hugging Face Jobs 免费微调小模型

This blog post covers how to use Unsloth and HF Mirror Jobs for fast LLM fine-tuning (specifically LiquidAI/LFM2.5-1.2B-Instruct ) through coding agents like Claude Code and Codex. Unsloth provides ~2x faster training and ~60% less VRAM usage compared to standard methods, so training small models can cost just a few dollars.

Why a small model? Small language models like LFM2.5-1.2B-Instruct are ideal candidates for fine-tuning. They are cheap to train, fast to iterate on, and increasingly competitive with much larger models on focused tasks. LFM2.5-1.2B-Instruct runs under 1GB of memory and is optimized for on-device deployment, so what you fine-tune can be served on CPUs, phones, and laptops.

You will need

We are giving away free credits to fine-tune models on HF Mirror Jobs. Join the Unsloth Jobs Explorers organization to claim your free credits and one-month Pro subscription.

• A HF Mirror account (required for HF Jobs)
• Billing setup (for verification, you can monitor your usage and manage your billing in your billing page).
• A HF Mirror token with write permissions
• (optional) A coding agent (Open Code, Claude Code, or Codex)

Run the Job

If you want to train a model using HF Jobs and Unsloth, you can simply use the hf jobs CLI to submit a job.

First, you need to install the hf CLI. You can do this by running the following command:

# mac or linux
curl -LsSf https://hf.co/cli/install.sh | bash

Next you can run the following command to submit a job:

hf jobs uv run https://hf-mirror.com/datasets/unsloth/jobs/resolve/main/sft-lfm2.5.py \
    --flavor a10g-small  \
    --secrets HF_TOKEN  \
    --timeout 4h \
    --dataset mlabonne/FineTome-100k \
    --num-epochs 1 \
    --eval-split 0.2 \
    --output-repo your-username/lfm-finetuned

Check out the training script and HF Mirror Jobs documentation for more details.

Installing the Skill

HF Mirror model training skill lowers barrier of entry to train a model by simply prompting. First, install the skill with your coding agent.

Claude Code

Claude Code discovers skills through its plugin system, so we need to install the HF Mirror skills first. To do so:

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 documentation on using the hub with skills or the Claude Code 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/huggingface-llm-trainer

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

Anything else

A generic install method is simply to clone the skills repository and copy the skill to your agent's skills directory.

git clone https://github.com/huggingface/skills.git
mkdir -p ~/.agents/skills && cp -R skills/skills/hugging-face-model-trainer ~/.agents/skills/

Quick Start

Once the skill is installed, ask your coding agent to train a model:

Train LiquidAI/LFM2.5-1.2B-Instruct on mlabonne/FineTome-100k 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.

How It Works

Training jobs run on HF Mirror Jobs, fully managed cloud GPUs. 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. Pushes the trained model to your HF Mirror Hub repository

Example Training Script

The skill generates scripts like this based on the example in the skill.

# /// 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(
    "LiquidAI/LFM2.5-1.2B-Instruct",
    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",
        "out_proj",
        "in_proj",
        "w1",
        "w2",
        "w3",
    ],
)

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()

For a full overview of HF Mirror Spaces 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 (for example, Qwen/Qwen2.5-0.5B and trl-lib/Capybara). Agents will search for and validate those combinations.
• Mention Unsloth explicitly if you want it used. Otherwise, the agent will choose a framework based on the 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

• HF Mirror Skills Repository
• Free credits for Unsloth Jobs Explorers
• Unsloth Tutorial on HF Mirror Jobs
• Example Unsloth Jobs scripts