Sovraine/prompt-injection-onnx

Sovraine Prompt Injection Detection (ONNX)

ONNX export of protectai/deberta-v3-base-prompt-injection-v2 for use with Sovraine Guard.

Model Details

• Architecture: DeBERTa-v3-base fine-tuned for prompt injection detection
• Parameters: 184 million
• Format: ONNX (runs on onnxruntime, no PyTorch required)
• Task: Binary classification (safe vs. injection)
• Language: English
• License: Apache 2.0

Usage

import onnxruntime as ort
from tokenizers import Tokenizer

session = ort.InferenceSession("model.onnx")
tokenizer = Tokenizer.from_file("tokenizer.json")
tokenizer.enable_padding(length=512)
tokenizer.enable_truncation(max_length=512)

enc = tokenizer.encode("your text here")
outputs = session.run(None, {
    "input_ids": [enc.ids],
    "attention_mask": [enc.attention_mask],
})
# outputs[0] shape: (1, 2) — [safe_score, injection_score]

Attribution

Derived from protectai/deberta-v3-base-prompt-injection-v2 (Apache 2.0). ONNX conversion by llmware (int4 quantized). Redistributed by SOVRAINE PTE.LTD. No fine-tuning applied.

Files

File	Description
model.onnx	ONNX model weights (int4 quantized)
tokenizer.json	HuggingFace tokenizer
spm.model	SentencePiece model
config.json	Model configuration