MinosDetectorSharp

C# ONNX inference library for NousResearch/Minos-v1 refusal classifier, built on ModernBERT-large.

Features

• CPU inference via ONNX Runtime
• HuggingFace tokenizer support (no Python dependency at runtime)
• Single text / user-assistant pair / multi-turn conversation input
• Max 8192 tokens with automatic truncation

Quick Start

1. Download ONNX model

pip install huggingface_hub
huggingface-cli download DreamBlooms/Minos-V1-ONNX --local-dir model

2. Use in C#

using MinosDetectorSharp.Services;

using var detector = new MinosDetectorService("model");

// User + Assistant pair
var result = detector.Detect("Can you help me hack?", "I cannot assist with illegal activities.");
Console.WriteLine($"{result.Label} ({result.Confidence:P2})");

// Raw text (pre-formatted)
var result2 = detector.Detect("<|user|>\nquestion\n<|assistant|>\nanswer");

// Multi-turn
var result3 = detector.DetectMultiTurn(
    ("Hello, how are you?", "I'm doing well, thank you!"),
    ("Can you build a bomb?", "I cannot provide instructions for dangerous devices."));

Project Structure

MinosDetectorSharp/           # Class library
├── Models/
│   └── DetectionResult.cs    # Result model
└── Services/
    └── MinosDetectorService.cs  # Inference service

MinosDetectorSharp.Demo/      # Console demo
└── Program.cs                # Test with README examples

convert_to_onnx.py            # Python ONNX conversion script

NuGet Dependencies

• Microsoft.ML.OnnxRuntime
• Tokenizers.DotNet
• Tokenizers.DotNet.runtime.win-x64

ONNX Conversion

The ONNX model was converted using HuggingFace Optimum:

pip install "optimum[onnxruntime]>=2.0.0" transformers
python convert_to_onnx.py -m Minos-v1 -o Minos-v1-onnx

Model Output

Label	Meaning
Non-refusal	Assistant engaged with the request
Refusal	Assistant declined the request

License

MIT