Spaces:
Sleeping
Sleeping
Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,187 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import gradio as gr
|
2 |
+
from huggingface_hub import InferenceClient
|
3 |
+
from datasets import load_dataset
|
4 |
+
import torch
|
5 |
+
from transformers import pipeline
|
6 |
+
|
7 |
+
class ContentFilter:
|
8 |
+
def __init__(self):
|
9 |
+
# Initialize toxic content detection model
|
10 |
+
self.toxicity_classifier = pipeline(
|
11 |
+
'text-classification',
|
12 |
+
model='unitary/toxic-bert',
|
13 |
+
return_all_scores=True
|
14 |
+
)
|
15 |
+
|
16 |
+
# Keyword blacklist
|
17 |
+
self.blacklist = [
|
18 |
+
'hate', 'discriminate', 'violent',
|
19 |
+
'offensive', 'inappropriate', 'racist',
|
20 |
+
'sexist', 'homophobic', 'transphobic'
|
21 |
+
]
|
22 |
+
|
23 |
+
def filter_toxicity(self, text, toxicity_threshold=0.7):
|
24 |
+
"""
|
25 |
+
Detect toxic content using pre-trained model
|
26 |
+
|
27 |
+
Args:
|
28 |
+
text (str): Input text to check
|
29 |
+
toxicity_threshold (float): Threshold for filtering
|
30 |
+
|
31 |
+
Returns:
|
32 |
+
dict: Filtering results
|
33 |
+
"""
|
34 |
+
results = self.toxicity_classifier(text)[0]
|
35 |
+
|
36 |
+
# Convert results to dictionary
|
37 |
+
toxicity_scores = {
|
38 |
+
result['label']: result['score']
|
39 |
+
for result in results
|
40 |
+
}
|
41 |
+
|
42 |
+
# Check if any toxic category exceeds threshold
|
43 |
+
is_toxic = any(
|
44 |
+
score > toxicity_threshold
|
45 |
+
for score in toxicity_scores.values()
|
46 |
+
)
|
47 |
+
|
48 |
+
return {
|
49 |
+
'is_toxic': is_toxic,
|
50 |
+
'toxicity_scores': toxicity_scores
|
51 |
+
}
|
52 |
+
|
53 |
+
def filter_keywords(self, text):
|
54 |
+
"""
|
55 |
+
Check text against keyword blacklist
|
56 |
+
|
57 |
+
Args:
|
58 |
+
text (str): Input text to check
|
59 |
+
|
60 |
+
Returns:
|
61 |
+
list: Matched blacklisted keywords
|
62 |
+
"""
|
63 |
+
matched_keywords = [
|
64 |
+
keyword for keyword in self.blacklist
|
65 |
+
if keyword.lower() in text.lower()
|
66 |
+
]
|
67 |
+
|
68 |
+
return matched_keywords
|
69 |
+
|
70 |
+
def comprehensive_filter(self, text):
|
71 |
+
"""
|
72 |
+
Perform comprehensive content filtering
|
73 |
+
|
74 |
+
Args:
|
75 |
+
text (str): Input text to filter
|
76 |
+
|
77 |
+
Returns:
|
78 |
+
dict: Comprehensive filtering results
|
79 |
+
"""
|
80 |
+
# Toxicity model filtering
|
81 |
+
toxicity_results = self.filter_toxicity(text)
|
82 |
+
|
83 |
+
# Keyword blacklist filtering
|
84 |
+
blacklisted_keywords = self.filter_keywords(text)
|
85 |
+
|
86 |
+
# Combine results
|
87 |
+
return {
|
88 |
+
'toxicity': toxicity_results,
|
89 |
+
'blacklisted_keywords': blacklisted_keywords,
|
90 |
+
'is_safe': not toxicity_results['is_toxic'] and len(blacklisted_keywords) == 0
|
91 |
+
}
|
92 |
+
|
93 |
+
# Initialize content filter
|
94 |
+
content_filter = ContentFilter()
|
95 |
+
|
96 |
+
# Initialize Hugging Face client
|
97 |
+
client = InferenceClient("HuggingFaceH4/zephyr-7b-beta")
|
98 |
+
|
99 |
+
# Load dataset (optional)
|
100 |
+
dataset = load_dataset("JustKiddo/KiddosVault")
|
101 |
+
|
102 |
+
def respond(
|
103 |
+
message,
|
104 |
+
history: list[tuple[str, str]],
|
105 |
+
system_message,
|
106 |
+
max_tokens,
|
107 |
+
temperature,
|
108 |
+
top_p
|
109 |
+
):
|
110 |
+
# First, filter the incoming user message
|
111 |
+
message_filter_result = content_filter.comprehensive_filter(message)
|
112 |
+
|
113 |
+
# If message is not safe, return a warning
|
114 |
+
if not message_filter_result['is_safe']:
|
115 |
+
toxicity_details = message_filter_result['toxicity']['toxicity_scores']
|
116 |
+
blacklisted_keywords = message_filter_result['blacklisted_keywords']
|
117 |
+
|
118 |
+
warning_message = "Message flagged for inappropriate content. "
|
119 |
+
warning_message += "Detected issues: "
|
120 |
+
|
121 |
+
# Add toxicity details
|
122 |
+
for category, score in toxicity_details.items():
|
123 |
+
if score > 0.7:
|
124 |
+
warning_message += f"{category} (Score: {score:.2f}), "
|
125 |
+
|
126 |
+
# Add blacklisted keywords
|
127 |
+
if blacklisted_keywords:
|
128 |
+
warning_message += f"Blacklisted keywords: {', '.join(blacklisted_keywords)}"
|
129 |
+
|
130 |
+
return warning_message
|
131 |
+
|
132 |
+
# Prepare messages for chat completion
|
133 |
+
messages = [{"role": "system", "content": system_message}]
|
134 |
+
for val in history:
|
135 |
+
if val[0]:
|
136 |
+
messages.append({"role": "user", "content": val[0]})
|
137 |
+
if val[1]:
|
138 |
+
messages.append({"role": "assistant", "content": val[1]})
|
139 |
+
messages.append({"role": "user", "content": message})
|
140 |
+
|
141 |
+
# Generate response
|
142 |
+
response = ""
|
143 |
+
for message in client.chat_completion(
|
144 |
+
messages,
|
145 |
+
max_tokens=max_tokens,
|
146 |
+
stream=True,
|
147 |
+
temperature=temperature,
|
148 |
+
top_p=top_p
|
149 |
+
):
|
150 |
+
token = message.choices[0].delta.content
|
151 |
+
response += token
|
152 |
+
yield response
|
153 |
+
|
154 |
+
# Create Gradio interface
|
155 |
+
demo = gr.ChatInterface(
|
156 |
+
respond,
|
157 |
+
additional_inputs=[
|
158 |
+
gr.Textbox(
|
159 |
+
value="You are a professional and friendly therapist specialized on LGBT+ issues",
|
160 |
+
label="System message"
|
161 |
+
),
|
162 |
+
gr.Slider(
|
163 |
+
minimum=1,
|
164 |
+
maximum=6144,
|
165 |
+
value=6144,
|
166 |
+
step=1,
|
167 |
+
label="Max new tokens"
|
168 |
+
),
|
169 |
+
gr.Slider(
|
170 |
+
minimum=0.1,
|
171 |
+
maximum=4.0,
|
172 |
+
value=1,
|
173 |
+
step=0.1,
|
174 |
+
label="Temperature"
|
175 |
+
),
|
176 |
+
gr.Slider(
|
177 |
+
minimum=0.1,
|
178 |
+
maximum=1.0,
|
179 |
+
value=0.95,
|
180 |
+
step=0.05,
|
181 |
+
label="Top-p (nucleus sampling)"
|
182 |
+
),
|
183 |
+
]
|
184 |
+
)
|
185 |
+
|
186 |
+
if __name__ == "__main__":
|
187 |
+
demo.launch(debug=True)
|