Prompt Injection Guard — Classifier v1
Model Description
DeBERTa-v3-base fine-tuned for binary prompt injection detection.
Classifies user inputs to LLM applications as benign or injection attempt.
Built as part of a 35-day portfolio project targeting the Anthropic Safeguards ML/Research Engineer role.
Training Data
Four public datasets unified into a single DuckDB schema after MinHash deduplication: 11,690 examples total.
Train/val/test split: 70/15/15, stratified by attack category.
Training Details
- Base model: microsoft/deberta-v3-base
- Epochs: 3, Learning rate: 2e-5, Batch size: 16
- Training time: ~28 minutes on Colab T4
Evaluation Results (Held-Out Test Set, n=1,754)
| Metric | Value |
|---|
| Macro F1 | 0.9957 |
| 95% Bootstrap CI | (0.9924, 0.9982) |
| Accuracy | 1.00 |
| Injection Recall | 1.00 |
| Injection Precision | 0.99 |
Per-Category F1
| Category | F1 | n |
|---|
| direct | 1.0000 | 92 |
| jailbreak | 1.0000 | 26 |
| system_prompt_leak | 1.0000 | 12 |
| unknown | 0.9962 | 1566 |
| role_play | 0.9644 | 57 |
Repository
