astoreyai/nspl-deberta-injection-v1

NSPL DeBERTa Prompt Injection Detector v1

Fine-tuned protectai/deberta-v3-base-prompt-injection-v2 on deepset/prompt-injections for improved recall on this benchmark.

Performance

Metric	Pretrained	Fine-tuned (this model)
Accuracy	67.2%	89.7%
Precision	100.0%	98.0%
Recall	36.7%	81.7%
F1	53.7%	89.1%

Cross-dataset (Ensemble with this model)

Dataset	Recall	F1
deepset test (n=116)	88.3%	92.2%
Jailbreak prompts (n=79)	100%	100%
jackhhao (n=1044)	98.7%	80.5%
xTRam1 (n=2000)	96.2%	81.3%
Lakera CTF (n=1000)	76.6%	86.8%

Usage

from nspl.guards.classifier import ClassifierDetector

detector = ClassifierDetector(
    model_name="astoreyai/nspl-deberta-injection-v1",
    threshold=0.5,
)
result = detector.detect("Ignore all previous instructions")
print(result.triggered)  # True
print(result.score)      # 0.98

Or standalone with transformers:

from transformers import pipeline

classifier = pipeline("text-classification", model="astoreyai/nspl-deberta-injection-v1")
result = classifier("Ignore all previous instructions")
# [{'label': 'INJECTION', 'score': 0.98}]

Training

• Base model: protectai/deberta-v3-base-prompt-injection-v2
• Dataset: deepset/prompt-injections (546 train, 116 test)
• Epochs: 3
• Learning rate: 2e-5
• Batch size: 16
• Training time: 56 seconds on CPU
• Framework: Hugging Face Transformers + Trainer

Part of NSPL

This model is part of NSPL (Non-Stochastic Protection Layer), a Python framework for deterministic LLM safety verification.

Citation

@software{storey2026nspl,
  author = {Storey, Aaron and McCardle, John},
  title = {NSPL: Non-Stochastic Protection Layer for Agentic AI},
  year = {2026},
  url = {https://github.com/astoreyai/nspl}
}