ahczhg/gpt-oss-safeguard-20b-demo

GPT-OSS-Safeguard-20B Demo

This repository demonstrates the usage of OpenAI's GPT-OSS-Safeguard-20B model for text generation tasks.

Overview

GPT-OSS-Safeguard-20B is a large language model designed for safe and responsible text generation. This project provides a simple implementation using the Hugging Face Transformers library.

Model Information

• Model: openai/gpt-oss-safeguard-20b
• Task: Text Generation
• Framework: Transformers (Hugging Face)

Features

• GPU-accelerated inference using CUDA
• Simple pipeline-based implementation
• Conversational text generation
• Easy-to-use interface for creative writing tasks

Requirements

• Python 3.8+
• CUDA-compatible GPU (recommended)
• Transformers library
• PyTorch

Installation

pip install -U transformers torch

For GPU support, ensure you have CUDA installed and configured properly.

Usage

The notebook demonstrates basic usage:

from transformers import pipeline

# Initialize the pipeline
pipe = pipeline("text-generation", model="openai/gpt-oss-safeguard-20b")

# Create a conversation
messages = [
    {"role": "user", "content": "Can you write a short story about a cat?"},
]

# Generate text
output = pipe(messages)
print(output[0]['generated_text'])

Files

• gpt-oss-safeguard-20b.ipynb: Jupyter notebook with example implementation

Notes

• The model uses MXFP4 quantization which requires Triton. If not available, it will default to bf16 dequantization.
• First run will download the model files (~13.76 GB total).
• Inference requires significant GPU memory.

Example Output

The model generates coherent and creative text based on the input prompts. In the example, when asked to write a short story about a cat, the model analyzes the request and begins crafting a narrative with character development and setting details.

Contributing

Feel free to open issues or submit pull requests for improvements.

License

MIT License - See LICENSE file for details.

Acknowledgments

• Model by OpenAI
• Built with Hugging Face Transformers