metadata

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

Dataset Classification with vLLM

Efficient text classification for Hugging Face datasets using vLLM with structured outputs. This script provides GPU-accelerated classification with guaranteed valid outputs through guided decoding.

🚀 Quick Start

# 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: This script uses vLLM for efficient inference
Python 3.10+
UV (will handle all dependencies automatically)
vLLM >= 0.6.6 (for guided decoding support)

🎯 Features

Guaranteed valid outputs using vLLM's guided decoding with outlines
Zero-shot classification with structured generation
GPU-optimized with vLLM's automatic batching for maximum efficiency
Robust text handling with preprocessing and validation
Three prompt styles for different use cases
Automatic progress tracking and detailed statistics
Direct Hub integration - read and write datasets seamlessly

💻 Usage

Basic Classification

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)
--prompt-style: Choose from simple, detailed, or reasoning (default: simple)
--split: Dataset split to process (default: train)
--max-samples: Limit samples for testing
--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)

Prompt Styles

simple: Direct classification prompt
detailed: Emphasizes exact category matching
reasoning: Includes brief analysis before classification

All styles benefit from structured output guarantees - the model can only output valid labels!

📊 Examples

Sentiment Analysis

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

Support Ticket Classification

uv run classify-dataset.py \
  --input-dataset user/support-tickets \
  --column content \
  --labels "bug,feature_request,question,other" \
  --output-dataset user/tickets-classified \
  --prompt-style reasoning

News Categorization

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

🚀 Running on HF Jobs

This script is optimized for Hugging Face Jobs (requires Pro subscription or Team/Enterprise organization):

# Run on L4 GPU with vLLM image
hf jobs uv run \
  --flavor l4x1 \
  --image vllm/vllm-openai:latest \
  classify-dataset.py \
  --input-dataset stanfordnlp/imdb \
  --column text \
  --labels "positive,negative" \
  --output-dataset user/imdb-classified

# Run on A10 GPU with custom model
hf jobs uv run \
  --flavor a10g-large \
  --image vllm/vllm-openai:latest \
  classify-dataset.py \
  --input-dataset user/reviews \
  --column review_text \
  --labels "1,2,3,4,5" \
  --output-dataset user/reviews-rated \
  --model mistralai/Mistral-7B-Instruct-v0.3 \
  --prompt-style detailed

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

Using Different Models

The default model is SmolLM3-3B, but you can use any instruction-tuned model:

# 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:

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

📈 Performance

SmolLM3-3B: ~50-100 texts/second on A10
7B models: ~20-50 texts/second on A10
vLLM automatically optimizes batching for best throughput

🤝 How It Works

vLLM: Provides efficient GPU batch inference
Guided Decoding: Uses outlines to guarantee valid label outputs
Structured Generation: Constrains model outputs to exact label choices
UV: Handles all dependencies automatically

The script loads your dataset, preprocesses texts, classifies each one using guided decoding to ensure only valid labels are generated, 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

📝 License

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