README.md

---
viewer: false
tags: [uv-script, classification, vllm, structured-outputs, gpu-required]
---

# Dataset Classification Script

GPU-accelerated text classification for Hugging Face datasets with guaranteed valid outputs through structured generation.

## 🚀 Quick Start

```bash
# Classify IMDB reviews
uv run classify-dataset.py \
  --input-dataset stanfordnlp/imdb \
  --column text \
  --labels "positive,negative" \
  --output-dataset user/imdb-classified
```

That's it! No installation, no setup - just `uv run`.

## 📋 Requirements

- **GPU Required**: Uses GPU-accelerated inference
- Python 3.10+
- UV (will handle all dependencies automatically)
- vLLM >= 0.6.6

## 🎯 Features

- **Guaranteed valid outputs** using structured generation with guided decoding
- **Zero-shot classification** without training data required
- **GPU-optimized** for maximum throughput and efficiency
- **Default model**: HuggingFaceTB/SmolLM3-3B (fast 3B model with thinking capabilities)
- **Robust text handling** with preprocessing and validation
- **Automatic progress tracking** and detailed statistics
- **Direct Hub integration** - read and write datasets seamlessly
- **Label descriptions** support for providing context to improve accuracy
- **Reasoning mode** for interpretable classifications with thinking traces
- **JSON output parsing** for reliable extraction from reasoning mode
- **Optimized batching** with vLLM's automatic batch processing
- **Multiple guided backends** - supports outlines, xgrammar, and more

## 💻 Usage

### Basic Classification

```bash
uv run classify-dataset.py \
  --input-dataset <dataset-id> \
  --column <text-column> \
  --labels <comma-separated-labels> \
  --output-dataset <output-id>
```

### Arguments

**Required:**

- `--input-dataset`: Hugging Face dataset ID (e.g., `stanfordnlp/imdb`, `user/my-dataset`)
- `--column`: Name of the text column to classify
- `--labels`: Comma-separated classification labels (e.g., `"spam,ham"`)
- `--output-dataset`: Where to save the classified dataset

**Optional:**

- `--model`: Model to use (default: **`HuggingFaceTB/SmolLM3-3B`** - a fast 3B parameter model)
- `--label-descriptions`: Provide descriptions for each label to improve classification accuracy
- `--enable-reasoning`: Enable reasoning mode with thinking traces (adds reasoning column)
- `--split`: Dataset split to process (default: `train`)
- `--max-samples`: Limit samples for testing
- `--shuffle`: Shuffle dataset before selecting samples (useful for random sampling)
- `--shuffle-seed`: Random seed for shuffling (default: 42)
- `--temperature`: Generation temperature (default: 0.1)
- `--guided-backend`: Backend for guided decoding (default: `outlines`)
- `--hf-token`: Hugging Face token (or use `HF_TOKEN` env var)

### Label Descriptions

Provide context for your labels to improve classification accuracy:

```bash
uv run classify-dataset.py \
  --input-dataset user/support-tickets \
  --column content \
  --labels "bug,feature,question,other" \
  --label-descriptions "bug:something is broken,feature:request for new functionality,question:asking for help,other:anything else" \
  --output-dataset user/tickets-classified
```

The model uses these descriptions to better understand what each label represents, leading to more accurate classifications.

### Reasoning Mode

Enable thinking traces for interpretable classifications:

```bash
uv run classify-dataset.py \
  --input-dataset stanfordnlp/imdb \
  --column text \
  --labels "positive,negative,neutral" \
  --enable-reasoning \
  --output-dataset user/imdb-with-reasoning
```

When `--enable-reasoning` is used:
- The model generates step-by-step reasoning using SmolLM3's thinking capabilities
- Output includes three columns: `classification`, `reasoning`, and `parsing_success`
- Final answer must be in JSON format: `{"label": "chosen_label"}`
- Useful for understanding complex classification decisions
- Trade-off: Slower but more interpretable

## 📊 Examples

### Sentiment Analysis

```bash
uv run classify-dataset.py \
  --input-dataset stanfordnlp/imdb \
  --column text \
  --labels "positive,negative" \
  --output-dataset user/imdb-sentiment
```

### Support Ticket Classification

```bash
uv run classify-dataset.py \
  --input-dataset user/support-tickets \
  --column content \
  --labels "bug,feature_request,question,other" \
  --label-descriptions "bug:code or product not working as expected,feature_request:asking for new functionality,question:seeking help or clarification,other:general comments or feedback" \
  --output-dataset user/tickets-classified
```

### News Categorization

```bash
uv run classify-dataset.py \
  --input-dataset ag_news \
  --column text \
  --labels "world,sports,business,tech" \
  --output-dataset user/ag-news-categorized \
  --model meta-llama/Llama-3.2-3B-Instruct
```

### Complex Classification with Reasoning

```bash
uv run classify-dataset.py \
  --input-dataset user/customer-feedback \
  --column text \
  --labels "very_positive,positive,neutral,negative,very_negative" \
  --label-descriptions "very_positive:extremely satisfied,positive:generally satisfied,neutral:mixed feelings,negative:dissatisfied,very_negative:extremely dissatisfied" \
  --enable-reasoning \
  --output-dataset user/feedback-analyzed
```

This combines label descriptions with reasoning mode for maximum interpretability.

### ArXiv ML Research Classification

Classify academic papers into machine learning research areas:

```bash
# Fast classification with random sampling
uv run classify-dataset.py \
  --input-dataset librarian-bots/arxiv-metadata-snapshot \
  --column abstract \
  --labels "llm,computer_vision,reinforcement_learning,optimization,theory,other" \
  --label-descriptions "llm:language models and NLP,computer_vision:image and video processing,reinforcement_learning:RL and decision making,optimization:training and efficiency,theory:theoretical ML foundations,other:other ML topics" \
  --output-dataset user/arxiv-ml-classified \
  --split "train[:10000]" \
  --max-samples 100 \
  --shuffle

# With reasoning for nuanced classification
uv run classify-dataset.py \
  --input-dataset librarian-bots/arxiv-metadata-snapshot \
  --column abstract \
  --labels "multimodal,agents,reasoning,safety,efficiency" \
  --label-descriptions "multimodal:vision-language and cross-modal models,agents:autonomous agents and tool use,reasoning:reasoning and planning systems,safety:alignment and safety research,efficiency:model optimization and deployment" \
  --enable-reasoning \
  --output-dataset user/arxiv-frontier-research \
  --split "train[:1000]" \
  --max-samples 50
```

The reasoning mode is particularly valuable for academic abstracts where papers often span multiple topics and require careful analysis to determine the primary focus.

## 🚀 Running on HF Jobs

Optimized for [Hugging Face Jobs](https://huggingface.co/docs/hub/spaces-gpu-jobs) (requires Pro subscription or Team/Enterprise organization):
```bash
# Run on L4 GPU with vLLM image
hf jobs uv run \
  --flavor l4x1 \
  --image vllm/vllm-openai:latest \
  https://huggingface.co/datasets/uv-scripts/classification/raw/main/classify-dataset.py \
  --input-dataset stanfordnlp/imdb \
  --column text \
  --labels "positive,negative" \
  --output-dataset user/imdb-classified
```

### GPU Flavors
- `t4-small`: Budget option for smaller models
- `l4x1`: Good balance for 7B models
- `a10g-small`: Fast inference for 3B models
- `a10g-large`: More memory for larger models
- `a100-large`: Maximum performance

## 🔧 Advanced Usage

### Random Sampling

When working with ordered datasets, use `--shuffle` with `--max-samples` to get a representative sample:

```bash
# Get 50 random reviews instead of the first 50
uv run classify-dataset.py \
  --input-dataset stanfordnlp/imdb \
  --column text \
  --labels "positive,negative" \
  --output-dataset user/imdb-sample \
  --max-samples 50 \
  --shuffle \
  --shuffle-seed 123  # For reproducibility
```

This is especially important for:
- Chronologically ordered datasets (news, papers, social media)
- Pre-sorted datasets (by rating, category, etc.)
- Testing on diverse samples before processing the full dataset

### Using Different Models

By default, this script uses **HuggingFaceTB/SmolLM3-3B** - a fast, efficient 3B parameter model that's perfect for most classification tasks. You can easily use any other instruction-tuned model:

```bash
# Larger model for complex classification
uv run classify-dataset.py \
  --input-dataset user/legal-docs \
  --column text \
  --labels "contract,patent,brief,memo,other" \
  --output-dataset user/legal-classified \
  --model Qwen/Qwen2.5-7B-Instruct
```

### Large Datasets

vLLM automatically handles batching for optimal performance. For very large datasets, it will process efficiently without manual intervention:

```bash
uv run classify-dataset.py \
  --input-dataset user/huge-dataset \
  --column text \
  --labels "A,B,C" \
  --output-dataset user/huge-classified
```

## 📈 Performance

- **SmolLM3-3B (default)**: ~50-100 texts/second on A10
- **7B models**: ~20-50 texts/second on A10
- vLLM automatically optimizes batching for best throughput
- Performance scales with GPU memory and compute capability

## 🤝 How It Works

1. **vLLM**: Provides efficient GPU batch inference with automatic batching
2. **Guided Decoding**: Uses outlines backend to guarantee valid label outputs
3. **Structured Generation**: Constrains model outputs to exact label choices
4. **UV**: Handles all dependencies automatically

The script loads your dataset, preprocesses texts, classifies each one with guaranteed valid outputs, then saves the results as a new column in the output dataset.

## 🐛 Troubleshooting

### CUDA Not Available

This script requires a GPU. Run it on:

- A machine with NVIDIA GPU
- HF Jobs (recommended)
- Cloud GPU instances

### Out of Memory

- Use a smaller model
- Use a larger GPU (e.g., a100-large)

### Invalid/Skipped Texts

- Texts shorter than 3 characters are skipped
- Empty or None values are marked as invalid
- Very long texts are truncated to 4000 characters

### Classification Quality

- With guided decoding, outputs are guaranteed to be valid labels
- For better results, use clear and distinct label names
- Try the `reasoning` prompt style for complex classifications
- Use a larger model for nuanced tasks

### vLLM Version Issues

If you see `ImportError: cannot import name 'GuidedDecodingParams'`:

- Your vLLM version is too old (requires >= 0.6.6)
- The script specifies the correct version in its dependencies
- UV should automatically install the correct version

## 🔬 Advanced Workflows

For complex real-world workflows that integrate UV scripts with the Python HF Jobs API, see the [ArXiv ML Trends example](examples/arxiv-workflow/). This demonstrates:

- **Multi-stage pipelines**: Data preparation → GPU classification → Analysis
- **Python API orchestration**: Using `run_uv_job()` to manage GPU jobs programmatically
- **Production patterns**: Error handling, parallel execution, and incremental updates
- **Cost optimization**: Choosing appropriate compute resources for each task

```python
# Example: Submit a classification job via Python API
from huggingface_hub import run_uv_job

job = run_uv_job(
    script="https://huggingface.co/datasets/uv-scripts/classification/raw/main/classify-dataset.py",
    args=["--input-dataset", "my/dataset", "--labels", "A,B,C"],
    flavor="l4x1",
    image="vllm/vllm-openai:latest"
)
result = job.wait()
```

## 📝 License

This script is provided as-is for use with the UV Scripts organization.