README.md

---
viewer: false
tags:
- uv-script
- training
- vlm
- unsloth
- iconclass
- fine-tuning
---

# VLM Training with Unsloth

Fine-tune Vision-Language Models efficiently using [Unsloth](https://github.com/unslothai/unsloth) - get 2x faster training with lower memory usage!

## 🎨 Example: Iconclass VLM

This directory contains scripts for fine-tuning VLMs to generate [Iconclass](https://iconclass.org) metadata codes from artwork images. Iconclass is a hierarchical classification system used in art history and cultural heritage.

### What You'll Train

Given an artwork image, the model outputs structured JSON:

```json
{
  "iconclass-codes": ["25H213", "25H216", "25I"]
}
```

Where codes represent:
- `25H213`: river
- `25H216`: waterfall
- `25I`: city-view with man-made constructions

## 🚀 Quick Start

### Option 1: Run on HF Jobs (Recommended)

```bash
# Set your HF token
export HF_TOKEN=your_token_here

# Submit training job
python submit_training_job.py
```

That's it! Your model will train on cloud GPUs and automatically push to the Hub.

### Option 2: Run Locally (Requires GPU)

```bash
# Install UV (if not already installed)
curl -LsSf https://astral.sh/uv/install.sh | sh

# Run training
uv run iconclass-vlm-sft.py \
  --base-model Qwen/Qwen3-VL-8B-Instruct \
  --dataset davanstrien/iconclass-vlm-sft \
  --output-model your-username/iconclass-vlm
```

### Option 3: Quick Test (100 steps)

```bash
uv run iconclass-vlm-sft.py \
  --base-model Qwen/Qwen3-VL-8B-Instruct \
  --dataset davanstrien/iconclass-vlm-sft \
  --output-model your-username/iconclass-vlm-test \
  --max-steps 100
```

## 📋 Requirements

### For HF Jobs
- Hugging Face account with Jobs access
- HF token with write permissions

### For Local Training
- CUDA-capable GPU (A100 recommended, A10G works)
- 40GB+ VRAM for 8B models (with 4-bit quantization)
- Python 3.11+
- [UV](https://docs.astral.sh/uv/) installed

## 🎛️ Configuration

### Quick Config via Python Script

Edit `submit_training_job.py`:

```python
# Model and dataset
BASE_MODEL = "Qwen/Qwen3-VL-8B-Instruct"
DATASET = "davanstrien/iconclass-vlm-sft"
OUTPUT_MODEL = "your-username/iconclass-vlm"

# Training settings
BATCH_SIZE = 2
GRADIENT_ACCUMULATION = 8
LEARNING_RATE = 2e-5
MAX_STEPS = None  # Auto-calculate for 1 epoch

# LoRA settings
LORA_R = 16
LORA_ALPHA = 32

# GPU
GPU_FLAVOR = "a100-large"  # or "a100", "a10g-large"
```

### Full CLI Options

```bash
uv run iconclass-vlm-sft.py --help
```

Key arguments:

| Argument | Default | Description |
|----------|---------|-------------|
| `--base-model` | Required | Base VLM (e.g., Qwen/Qwen3-VL-8B-Instruct) |
| `--dataset` | Required | Training dataset on HF Hub |
| `--output-model` | Required | Where to push your model |
| `--lora-r` | 16 | LoRA rank (higher = more capacity) |
| `--lora-alpha` | 32 | LoRA alpha (usually 2×r) |
| `--learning-rate` | 2e-5 | Learning rate |
| `--batch-size` | 2 | Per-device batch size |
| `--gradient-accumulation` | 8 | Gradient accumulation steps |
| `--max-steps` | Auto | Total training steps |
| `--num-epochs` | 1.0 | Epochs (if max-steps not set) |

## 🏗️ Architecture

### What Makes This Fast?

1. **Unsloth Optimizations**: 2x faster training through:
   - Optimized CUDA kernels
   - Better memory management
   - Efficient gradient checkpointing

2. **4-bit Quantization**:
   - Loads model in 4-bit precision
   - Dramatically reduces VRAM usage
   - Minimal impact on quality with LoRA

3. **LoRA (Low-Rank Adaptation)**:
   - Only trains 0.1-1% of parameters
   - Much faster than full fine-tuning
   - Easy to merge back or share

### Training Flow

```
Dataset (HF Hub)
    ↓
FastVisionModel.from_pretrained (4-bit)
    ↓
Apply LoRA adapters
    ↓
SFTTrainer (Unsloth-optimized)
    ↓
Push to Hub with model card
```

## 📊 Expected Performance

### Training Time (Qwen3-VL-8B on A100)

| Dataset Size | Batch Config | Time | Cost (est.) |
|--------------|--------------|------|-------------|
| 44K samples | BS=2, GA=8 | ~4h | $16 |
| 10K samples | BS=2, GA=8 | ~1h | $4 |
| 1K samples | BS=2, GA=8 | ~10min | $0.70 |

*BS = Batch Size, GA = Gradient Accumulation*

### GPU Requirements

| Model Size | Min GPU | Recommended | VRAM Usage |
|------------|---------|-------------|------------|
| 3B-4B | A10G | A100 | ~20GB |
| 7B-8B | A100 | A100 | ~35GB |
| 13B+ | A100 (80GB) | A100 (80GB) | ~60GB |

## 🔍 Monitoring Your Job

### Via CLI

```bash
# Check status
hfjobs status your-job-id

# Stream logs
hfjobs logs your-job-id --follow

# List all jobs
hfjobs list
```

### Via Python

```python
from huggingface_hub import HfApi

api = HfApi()
job = api.get_job("your-job-id")

print(job.status)
print(job.logs())
```

### Via Web

Your job URL: `https://huggingface.co/jobs/your-username/your-job-id`

## 🎯 Using Your Fine-Tuned Model

```python
from unsloth import FastVisionModel
from PIL import Image

# Load your model
model, tokenizer = FastVisionModel.from_pretrained(
    model_name="your-username/iconclass-vlm",
    load_in_4bit=True,
    max_seq_length=2048,
)
FastVisionModel.for_inference(model)

# Prepare input
image = Image.open("artwork.jpg")
prompt = "Extract ICONCLASS labels for this image."

messages = [
    {
        "role": "user",
        "content": [
            {"type": "image"},
            {"type": "text", "text": prompt},
        ],
    }
]

# Apply chat template
inputs = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    return_tensors="pt",
).to("cuda")

# Generate
outputs = model.generate(
    **inputs,
    max_new_tokens=256,
    temperature=0.7,
    top_p=0.9,
)

response = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(response)
# {"iconclass-codes": ["31A235", "31A24(+1)", "61B(+54)"]}
```

## 📦 Files in This Directory

| File | Purpose |
|------|---------|
| `iconclass-vlm-sft.py` | Main training script (UV script) |
| `submit_training_job.py` | Helper to submit HF Jobs |
| `README.md` | This file |

## 🛠️ Troubleshooting

### Out of Memory?

Reduce batch size or increase gradient accumulation:
```bash
--batch-size 1 --gradient-accumulation 16
```

### Training Too Slow?

Increase batch size if you have VRAM:
```bash
--batch-size 4 --gradient-accumulation 4
```

### Model Not Learning?

Try adjusting learning rate:
```bash
--learning-rate 5e-5  # Higher
--learning-rate 1e-5  # Lower
```

Or increase LoRA rank:
```bash
--lora-r 32 --lora-alpha 64
```

### Jobs Failing?

Check logs:
```bash
hfjobs logs your-job-id
```

Common issues:
- HF_TOKEN not set correctly
- Output model repo doesn't exist (create it first)
- GPU out of memory (reduce batch size)

## 🔗 Related Resources

- **Unsloth**: https://github.com/unslothai/unsloth
- **Unsloth Docs**: https://docs.unsloth.ai/
- **TRL**: https://github.com/huggingface/trl
- **HF Jobs**: https://huggingface.co/docs/hub/spaces-sdks-jobs
- **UV**: https://docs.astral.sh/uv/
- **Iconclass**: https://iconclass.org
- **Blog Post**: https://danielvanstrien.xyz/posts/2025/iconclass-vlm-sft/

## 💡 Tips

1. **Start Small**: Test with `--max-steps 100` before full training
2. **Use Wandb**: Add `--report-to wandb` for better monitoring
3. **Save Often**: Use `--save-steps 50` for checkpoints
4. **Multiple GPUs**: Script automatically uses all available GPUs
5. **Resume Training**: Load from checkpoint with `--resume-from-checkpoint`

## 📝 Citation

If you use this training setup, please cite:

```bibtex
@misc{iconclass-vlm-training,
  author = {Daniel van Strien},
  title = {Efficient VLM Fine-tuning with Unsloth for Art History},
  year = {2025},
  publisher = {GitHub},
  howpublished = {\url{https://github.com/davanstrien/uv-scripts}}
}
```

---

Made with 🦥 [Unsloth](https://github.com/unslothai/unsloth) •
Powered by 🤖 [UV Scripts](https://huggingface.co/uv-scripts)