uv-scripts/unsloth-jobs

Unsloth + HF Jobs: Zero-Setup Efficient Training

Fine-tune LLMs and VLMs on cloud GPUs without any environment setup. Unsloth provides 2x faster training with 60% less VRAM, and HF Jobs gives you on-demand A100s that stream data directly from the Hub.

No Docker. No pip install. No CUDA setup. Just run.

What You Need

1. A Hugging Face account with a token
2. The HF CLI - install with:

   curl -LsSf https://hf.co/cli/install.sh | bash
   

3. A dataset on the Hub (see format requirements below)

Prepare Your Data

For VLM Fine-tuning (images + text)

Your dataset needs two columns: images and messages.

{
    "images": [<PIL.Image>],  # List of images
    "messages": [
        {
            "role": "user",
            "content": [
                {"type": "image"},
                {"type": "text", "text": "What's in this image?"}
            ]
        },
        {
            "role": "assistant",
            "content": [
                {"type": "text", "text": "A golden retriever playing fetch in a park."}
            ]
        }
    ]
}

See davanstrien/iconclass-vlm-sft for a working example.

For Continued Pretraining (text only)

Any dataset with a text column works:

{"text": "Your domain-specific text here..."}

Use --text-column if your column has a different name.

Step 1: Test Locally (Optional)

Make sure your dataset format is correct by running a quick local test:

# Check the script works (shows help)
uv run https://huggingface.co/datasets/uv-scripts/unsloth-jobs/raw/main/sft-qwen3-vl.py --help

Step 2: Run on HF Jobs

Fine-tune a Vision-Language Model

hf jobs uv run \
  https://huggingface.co/datasets/uv-scripts/unsloth-jobs/raw/main/sft-qwen3-vl.py \
  --flavor a100-large --secrets HF_TOKEN --timeout 4h \
  -- --dataset your-username/your-vlm-dataset \
     --num-epochs 1 \
     --eval-split 0.2 \
     --output-repo your-username/my-vlm

What this does:

• Spins up an A100 GPU on HF Jobs
• Downloads and installs all dependencies automatically
• Loads Qwen3-VL-8B with Unsloth optimizations
• Trains for 1 epoch, holding out 20% for evaluation
• Uploads your fine-tuned model to the Hub

Continued Pretraining on Domain Text

hf jobs uv run \
  https://huggingface.co/datasets/uv-scripts/unsloth-jobs/raw/main/continued-pretraining.py \
  --flavor a100-large --secrets HF_TOKEN \
  -- --dataset your-username/domain-corpus \
     --text-column content \
     --max-steps 1000 \
     --output-repo your-username/domain-llm

Step 3: Monitor Progress (Optional)

Add Trackio for real-time training metrics:

hf jobs uv run \
  https://huggingface.co/datasets/uv-scripts/unsloth-jobs/raw/main/sft-qwen3-vl.py \
  --flavor a100-large --secrets HF_TOKEN \
  -- --dataset your-username/dataset \
     --trackio-space your-username/trackio \
     --output-repo your-username/my-model

Available Scripts

Script	Base Model	Best For
sft-qwen3-vl.py	Qwen3-VL-8B	High-quality VLM fine-tuning
sft-gemma3-vlm.py	Gemma 3 4B	Lightweight/faster VLM tasks
continued-pretraining.py	Qwen3-0.6B	Domain adaptation, new languages

Common Options

Option	Description	Default
--dataset	HF dataset ID	required
--output-repo	Where to save trained model	required
--max-steps	Number of training steps	500
--num-epochs	Train for N epochs instead of steps	-
--eval-split	Fraction for evaluation (e.g., 0.2)	0 (disabled)
--batch-size	Per-device batch size	2
--learning-rate	Learning rate	2e-4
--trackio-space	HF Space for live monitoring	-

Run any script with --help to see all options:

uv run https://huggingface.co/datasets/uv-scripts/unsloth-jobs/raw/main/sft-qwen3-vl.py --help

Tips

• Start small: Use --max-steps 10 to verify everything works before a full run
• Use eval splits: --eval-split 0.2 helps detect overfitting
• Check costs: A100-large is ~$4/hr, estimate your training time first
• Streaming for large datasets: Add --streaming if your dataset is very large

How It Works

1. hf jobs uv run spins up an A100 GPU
2. UV reads dependencies from the script header and installs them
3. Unsloth loads the model with 4-bit quantization and LoRA
4. Training streams data directly from the Hub (fast!)
5. Your fine-tuned adapter uploads to the Hub automatically

Learn More

• HF Jobs Quickstart - Getting started with HF Jobs
• Unsloth Documentation - Training optimizations
• UV Scripts Guide - How inline dependencies work