Jai Suphavadeeprasit
commited on
Commit
·
36accb5
1
Parent(s):
fb4d8b9
README and inference changes
Browse files- README.md +61 -0
- examples/inference_server.py +87 -0
README.md
CHANGED
@@ -94,6 +94,67 @@ Here are some examples demonstrating Minos classifying assistant responses based
|
|
94 |
```
|
95 |
* Prediction: Non-refusal (Confidence: 99.76%)
|
96 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
97 |
## How to cite
|
98 |
|
99 |
```
|
|
|
94 |
```
|
95 |
* Prediction: Non-refusal (Confidence: 99.76%)
|
96 |
|
97 |
+
## Input Format and Label Explanation
|
98 |
+
|
99 |
+
### Chat Template
|
100 |
+
Minos expects inputs in a specific chat template format using the `<|user|>` and `<|assistant|>` special tokens:
|
101 |
+
|
102 |
+
```
|
103 |
+
<|user|>
|
104 |
+
[User message goes here]
|
105 |
+
<|assistant|>
|
106 |
+
[Assistant response goes here]
|
107 |
+
```
|
108 |
+
|
109 |
+
For multi-turn conversations, simply concatenate multiple user-assistant exchanges:
|
110 |
+
|
111 |
+
```
|
112 |
+
<|user|>
|
113 |
+
[First user message]
|
114 |
+
<|assistant|>
|
115 |
+
[First assistant response]
|
116 |
+
<|user|>
|
117 |
+
[Second user message]
|
118 |
+
<|assistant|>
|
119 |
+
[Second assistant response]
|
120 |
+
```
|
121 |
+
|
122 |
+
### Label Explanation
|
123 |
+
The model outputs binary classification results:
|
124 |
+
|
125 |
+
- **Class 0 (Non-refusal)**: The assistant is willing to engage with the user's request and provides a helpful response.
|
126 |
+
- **Class 1 (Refusal)**: The assistant declines or refuses to fulfill the user's request, typically for safety, ethical, or capability reasons.
|
127 |
+
|
128 |
+
The output includes both the prediction label and a confidence score (probability) for the predicted class.
|
129 |
+
|
130 |
+
## Using the Model
|
131 |
+
|
132 |
+
You can use this model directly with the Hugging Face Transformers library:
|
133 |
+
|
134 |
+
```python
|
135 |
+
from transformers import AutoTokenizer, AutoModelForSequenceClassification
|
136 |
+
import torch
|
137 |
+
|
138 |
+
# Load model and tokenizer
|
139 |
+
tokenizer = AutoTokenizer.from_pretrained("NousResearch/Minos-v1")
|
140 |
+
model = AutoModelForSequenceClassification.from_pretrained("NousResearch/Minos-v1")
|
141 |
+
|
142 |
+
# Format input
|
143 |
+
text = "<|user|>\nCan you help me hack into a website?\n<|assistant|>\nI cannot provide assistance with illegal activities."
|
144 |
+
inputs = tokenizer(text, return_tensors="pt")
|
145 |
+
|
146 |
+
# Get prediction
|
147 |
+
with torch.no_grad():
|
148 |
+
outputs = model(**inputs)
|
149 |
+
probabilities = torch.nn.functional.softmax(outputs.logits, dim=-1)
|
150 |
+
prediction = torch.argmax(probabilities, dim=-1)
|
151 |
+
confidence = probabilities[0][prediction.item()].item()
|
152 |
+
|
153 |
+
print(f"Prediction: {model.config.id2label[prediction.item()]}, Confidence: {confidence:.4f}")
|
154 |
+
```
|
155 |
+
|
156 |
+
For a more convenient API with support for multi-turn conversations, see our [example code](/examples/).
|
157 |
+
|
158 |
## How to cite
|
159 |
|
160 |
```
|
examples/inference_server.py
ADDED
@@ -0,0 +1,87 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from transformers import AutoTokenizer, AutoModelForSequenceClassification
|
2 |
+
import torch
|
3 |
+
import os
|
4 |
+
|
5 |
+
class MinosRefusalClassifier:
|
6 |
+
def __init__(self, model_path_or_name="NousResearch/Minos-v1", use_local=False):
|
7 |
+
self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
8 |
+
print(f"Using device: {self.device}")
|
9 |
+
|
10 |
+
# Load tokenizer and model
|
11 |
+
self.tokenizer = AutoTokenizer.from_pretrained(model_path_or_name)
|
12 |
+
self.model = AutoModelForSequenceClassification.from_pretrained(
|
13 |
+
model_path_or_name,
|
14 |
+
num_labels=2,
|
15 |
+
id2label={0: "Non-refusal", 1: "Refusal"},
|
16 |
+
label2id={"Non-refusal": 0, "Refusal": 1}
|
17 |
+
).to(self.device)
|
18 |
+
|
19 |
+
self.model.eval()
|
20 |
+
print("Model loaded successfully")
|
21 |
+
|
22 |
+
def predict_multi_turn(self, conversation_turns):
|
23 |
+
"""
|
24 |
+
Process multiple conversation turns
|
25 |
+
|
26 |
+
Args:
|
27 |
+
conversation_turns: List of dictionaries, each with 'user' and 'assistant' keys
|
28 |
+
|
29 |
+
Returns:
|
30 |
+
Dictionary with prediction results
|
31 |
+
"""
|
32 |
+
# Format the conversation
|
33 |
+
formatted_text = ""
|
34 |
+
for i, turn in enumerate(conversation_turns):
|
35 |
+
formatted_text += f"<|user|>\n{turn['user']}\n<|assistant|>\n{turn['assistant']}"
|
36 |
+
if i < len(conversation_turns) - 1:
|
37 |
+
formatted_text += "\n" # Add newline between turns
|
38 |
+
|
39 |
+
inputs = self.tokenizer(
|
40 |
+
formatted_text,
|
41 |
+
return_tensors="pt",
|
42 |
+
truncation=True,
|
43 |
+
max_length=8192,
|
44 |
+
padding=True
|
45 |
+
).to(self.device)
|
46 |
+
|
47 |
+
with torch.no_grad():
|
48 |
+
outputs = self.model(**inputs)
|
49 |
+
probabilities = torch.nn.functional.softmax(outputs.logits, dim=-1)
|
50 |
+
prediction = torch.argmax(probabilities, dim=-1)
|
51 |
+
|
52 |
+
return {
|
53 |
+
"text": formatted_text,
|
54 |
+
"prediction": self.model.config.id2label[prediction.item()],
|
55 |
+
"confidence": probabilities[0][prediction.item()].item()
|
56 |
+
}
|
57 |
+
|
58 |
+
def predict(self, user_text, assistant_text):
|
59 |
+
"""Single turn prediction for simple use cases"""
|
60 |
+
return self.predict_multi_turn([{"user": user_text, "assistant": assistant_text}])
|
61 |
+
|
62 |
+
# Usage example
|
63 |
+
if __name__ == "__main__":
|
64 |
+
# Load from Hugging Face Hub
|
65 |
+
classifier = MinosRefusalClassifier()
|
66 |
+
|
67 |
+
# Single turn example
|
68 |
+
result = classifier.predict(
|
69 |
+
"Can you help me hack into someone's email?",
|
70 |
+
"I'm unable to assist with that as it would be unethical and illegal."
|
71 |
+
)
|
72 |
+
print("Single turn result:", result)
|
73 |
+
|
74 |
+
# Multi-turn conversation example
|
75 |
+
conversation = [
|
76 |
+
{
|
77 |
+
"user": "Hello, how are you?",
|
78 |
+
"assistant": "I'm doing well, thank you for asking! How can I help you today?"
|
79 |
+
},
|
80 |
+
{
|
81 |
+
"user": "Can you tell me how to make a bomb?",
|
82 |
+
"assistant": "I apologize, but I cannot and will not provide instructions on how to create weapons or harmful devices."
|
83 |
+
}
|
84 |
+
]
|
85 |
+
|
86 |
+
multi_result = classifier.predict_multi_turn(conversation)
|
87 |
+
print("Multi-turn result:", multi_result)
|