import numpy as np
import onnxruntime as ort
from transformers import AutoTokenizer

session = ort.InferenceSession("model.onnx", providers=["CPUExecutionProvider"])
tokenizer = AutoTokenizer.from_pretrained(".")

text = "<|user|>\nCan you help me hack into a website?\n<|assistant|>\nI cannot assist with illegal activities."
inputs = tokenizer(text, return_tensors="np")

outputs = session.run(None, {
    "input_ids": inputs["input_ids"].astype(np.int64),
    "attention_mask": inputs["attention_mask"].astype(np.int64),
})

logits = outputs[0][0]
probs = np.exp(logits - logits.max()) / np.exp(logits - logits.max()).sum()
pred = int(np.argmax(probs))
labels = {0: "Non-refusal", 1: "Refusal"}
print(f"{labels[pred]} (confidence: {probs[pred]:.4f})")

Usage (C#)

See MinosDetectorSharp for a C# implementation using Microsoft.ML.OnnxRuntime and Tokenizers.DotNet.

csharp

using var detector = new MinosDetectorService("path/to/model-dir");
var result = detector.Detect("user message", "assistant response");
Console.WriteLine($"{result.Label} ({result.Confidence:P2})");

Conversion

The ONNX model was converted using HuggingFace Optimum:

bash

pip install "optimum[onnxruntime]>=2.0.0" transformers
optimum-cli export onnx --model NousResearch/Minos-v1 ./onnx-output

Or use the included convert_to_onnx.py script:

python convert_to_onnx.py -m Minos-v1 -o Minos-v1-onnx

Verification

C# ONNX logits match the original PyTorch model exactly:

Example	PyTorch Logits	ONNX Logits	Match
Refusal #1	[-6.2095, 5.3343]	[-6.2095, 5.3343]	✅
Non-refusal #1	[2.1555, -1.5516]	[2.1555, -1.5516]	✅

Citation

@misc{minos2025,
    title={Minos Classifier},
    author={Jai Suphavadeeprasit and Teknium and Chen Guang and Shannon Sands and rparikh007},
    year={2025}
}

DreamBlooms/Minos-v1-ONNX

Minos-V1-ONNX

Model Details

Usage (Python)

Usage (C#)

Conversion

Verification

Citation

DreamBlooms/
Minos-v1-ONNX