Project · 2025
PromptGuard
Fine-tuned DeBERTa on ~51k examples for prompt injection detection. Deployed to Hugging Face in 16-bit and 8-bit variants. 3.33ms CPU / 0.44ms GPU inference.
Overview
Fine-tuned DeBERTa on ~51k examples for prompt injection detection. Deployed to Hugging Face in 16-bit and 8-bit variants. 3.33ms CPU / 0.44ms GPU inference.
Problem
Prompt injection attempts can undermine applications built around language models.
Approach
Fine-tuned DeBERTa on approximately 51,000 examples for prompt-injection detection and published 16-bit and 8-bit models on Hugging Face.
Results
Project measurements report 3.33 ms CPU inference and 0.44 ms GPU inference.