README.md

---
viewer: false
tags:
- uv-script
- training
- unsloth
- streaming
- fine-tuning
- llm
---

# Streaming LLM Training with Unsloth

Train on massive datasets without downloading anything - data streams directly from the Hub.

## 🦥 Latin LLM Example

Teaches Qwen Latin using 1.47M texts from FineWeb-2, streamed directly from the Hub.

**Blog post:** [Train on Massive Datasets Without Downloading](https://danielvanstrien.xyz/posts/2026/hf-streaming-unsloth/train-massive-datasets-without-downloading.html)

### Quick Start

```bash
# Run on HF Jobs (recommended - 2x faster streaming)
hf jobs uv run latin-llm-streaming.py \
  --flavor a100-large \
  --timeout 2h \
  --secrets HF_TOKEN \
  -- \
  --max-steps 500 \
  --output-repo your-username/qwen-latin

# Run locally
uv run latin-llm-streaming.py \
  --max-steps 100 \
  --output-repo your-username/qwen-latin-test
```

### Why Streaming?

- **No disk space needed** - train on TB-scale datasets without downloading
- **Works everywhere** - Colab, Kaggle, HF Jobs
- **Any language** - FineWeb-2 has 90+ languages available

### Options

| Argument | Default | Description |
|----------|---------|-------------|
| `--base-model` | `unsloth/Qwen3-0.6B-Base-unsloth-bnb-4bit` | Base model |
| `--max-steps` | 500 | Training steps |
| `--batch-size` | 4 | Per-device batch size |
| `--gradient-accumulation` | 4 | Gradient accumulation steps |
| `--learning-rate` | 2e-4 | Learning rate |
| `--output-repo` | Required | Where to push model |
| `--wandb-project` | None | Wandb project for logging |

### Performance

| Environment | Speed | Why |
|-------------|-------|-----|
| Colab A100 | ~0.36 it/s | Network latency |
| HF Jobs A100 | ~0.74 it/s | Co-located compute |

Streaming is ~2x faster on HF Jobs because compute is co-located with the data.

---

## 🎨 VLM Streaming Fine-tuning (Qwen3-VL)

Fine-tune Vision Language Models with streaming datasets - ideal for large image-text datasets.

**Script:** `vlm-streaming-sft-unsloth-qwen.py`
**Default model:** `unsloth/Qwen3-VL-8B-Instruct-unsloth-bnb-4bit`
**Example dataset:** [`davanstrien/iconclass-vlm-sft`](https://huggingface.co/datasets/davanstrien/iconclass-vlm-sft)

> **Note:** This script uses pinned dependencies (`transformers==4.57.1`, `trl==0.22.2`) matching the [official Unsloth Qwen3-VL notebook](https://colab.research.google.com/github/unslothai/notebooks/blob/main/nb/Qwen3_VL_(7B)-Vision.ipynb) for maximum compatibility.

### Quick Start

```bash
# Run on HF Jobs (recommended)
hf jobs uv run \
  --flavor a100-large \
  --secrets HF_TOKEN \
  -- \
  https://huggingface.co/datasets/uv-scripts/training/raw/main/vlm-streaming-sft-unsloth-qwen.py \
  --max-steps 500 \
  --output-repo your-username/vlm-finetuned

# With Trackio monitoring dashboard
hf jobs uv run \
  --flavor a100-large \
  --secrets HF_TOKEN \
  -- \
  https://huggingface.co/datasets/uv-scripts/training/raw/main/vlm-streaming-sft-unsloth-qwen.py \
  --max-steps 500 \
  --output-repo your-username/vlm-finetuned \
  --trackio-space your-username/trackio
```

### Why Streaming for VLMs?

- **No disk space needed** - images stream directly from Hub
- **Works with massive datasets** - train on datasets larger than your storage
- **Memory efficient** - Unsloth uses ~60% less VRAM
- **2x faster** - Unsloth optimizations for Qwen3-VL

### Verified Performance

Tested on HF Jobs with A100-80GB:

| Setting | Value |
|---------|-------|
| Model | Qwen3-VL-8B (4-bit) |
| Dataset | iconclass-vlm-sft |
| Speed | ~3s/step |
| 50 steps | ~3 minutes |
| Starting loss | 4.3 |
| Final loss | ~0.85 |

### Options

| Argument | Default | Description |
|----------|---------|-------------|
| `--base-model` | `unsloth/Qwen3-VL-8B-Instruct-unsloth-bnb-4bit` | Base VLM model |
| `--dataset` | `davanstrien/iconclass-vlm-sft` | Dataset with images + messages |
| `--max-steps` | 500 | Training steps (required for streaming) |
| `--batch-size` | 2 | Per-device batch size |
| `--gradient-accumulation` | 4 | Gradient accumulation steps |
| `--learning-rate` | 2e-4 | Learning rate |
| `--lora-r` | 16 | LoRA rank |
| `--lora-alpha` | 16 | LoRA alpha (same as r per Unsloth notebook) |
| `--output-repo` | Required | Where to push model |
| `--trackio-space` | None | HF Space for Trackio dashboard |

### Dataset Format

The script works with **any dataset** that has `images` and `messages` columns in the standard VLM conversation format:

```python
{
    "images": [<PIL.Image>],  # Single image or list of images
    "messages": [
        {"role": "user", "content": [{"type": "image"}, {"type": "text", "text": "Describe this image"}]},
        {"role": "assistant", "content": [{"type": "text", "text": "The image shows..."}]}
    ]
}
```

**Compatible datasets:**
- [`davanstrien/iconclass-vlm-sft`](https://huggingface.co/datasets/davanstrien/iconclass-vlm-sft) - Art iconography classification
- Any dataset following the [Unsloth VLM format](https://docs.unsloth.ai/basics/vision-finetuning)

### Calculating Steps from Dataset Size

Since streaming datasets don't expose their length, use this formula:
```
steps = dataset_size / (batch_size * gradient_accumulation)
```

For example, with 10,000 samples, batch_size=2, gradient_accumulation=4:
```
steps = 10000 / (2 * 4) = 1250 steps for 1 epoch
```

---

## 🚀 Running on HF Jobs

```bash
# Basic usage
hf jobs uv run latin-llm-streaming.py --flavor a100-large --secrets HF_TOKEN

# With timeout for long training
hf jobs uv run latin-llm-streaming.py --flavor a100-large --timeout 2h --secrets HF_TOKEN

# Pass script arguments after --
hf jobs uv run latin-llm-streaming.py --flavor a100-large -- --max-steps 1000 --batch-size 8
```

### Available Flavors

- `a100-large` - 80GB VRAM (recommended)
- `a10g-large` - 24GB VRAM
- `t4-small` - 16GB VRAM

---

## 🔗 Resources

- [Unsloth](https://github.com/unslothai/unsloth) - 2x faster training
- [HF Jobs Docs](https://huggingface.co/docs/huggingface_hub/guides/jobs)
- [Datasets Streaming](https://huggingface.co/docs/datasets/stream)
- [Streaming Datasets Blog](https://huggingface.co/blog/streaming-datasets)

---

Made with 🦥 [Unsloth](https://github.com/unslothai/unsloth)