Spaces:
Running
Running
Commit
•
5d09640
1
Parent(s):
80f9197
first commit
Browse files
app.py
CHANGED
@@ -1,3 +1,59 @@
|
|
1 |
-
import gradio as gr
|
|
|
2 |
|
3 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
#import gradio as gr
|
2 |
+
#gr.load("models/walledai/walledguard-c").launch()
|
3 |
|
4 |
+
import streamlit as st
|
5 |
+
import torch
|
6 |
+
import torch.nn as nn
|
7 |
+
from transformers import AutoTokenizer, AutoModelForCausalLM
|
8 |
+
|
9 |
+
# Define the template
|
10 |
+
TEMPLATE = """[INST] Evaluate if the given text is harmful i.e. either it asks for unsafe information or it contains unsafe information.
|
11 |
+
<START TEXT>
|
12 |
+
{prompt}
|
13 |
+
<END TEXT>
|
14 |
+
Answer: [/INST]
|
15 |
+
"""
|
16 |
+
|
17 |
+
# Load the model and tokenizer
|
18 |
+
@st.cache_resource
|
19 |
+
def load_model():
|
20 |
+
model_name = "walledai/walledguard-c"
|
21 |
+
tokenizer = AutoTokenizer.from_pretrained(model_name)
|
22 |
+
model = AutoModelForCausalLM.from_pretrained(model_name)
|
23 |
+
return tokenizer, model
|
24 |
+
|
25 |
+
tokenizer, model = load_model()
|
26 |
+
|
27 |
+
# Streamlit app
|
28 |
+
st.title("Text Safety Evaluator")
|
29 |
+
|
30 |
+
# User input
|
31 |
+
user_input = st.text_area("Enter the text you want to evaluate:", height=100)
|
32 |
+
|
33 |
+
if st.button("Evaluate"):
|
34 |
+
if user_input:
|
35 |
+
# Prepare input
|
36 |
+
input_ids = tokenizer.encode(TEMPLATE.format(prompt=user_input), return_tensors="pt")
|
37 |
+
|
38 |
+
# Generate output
|
39 |
+
output = model.generate(input_ids=input_ids, max_new_tokens=20, pad_token_id=0)
|
40 |
+
|
41 |
+
# Decode output
|
42 |
+
prompt_len = input_ids.shape[-1]
|
43 |
+
output_decoded = tokenizer.decode(output[0][prompt_len:], skip_special_tokens=True)
|
44 |
+
|
45 |
+
# Determine prediction
|
46 |
+
prediction = 'unsafe' if 'unsafe' in output_decoded.lower() else 'safe'
|
47 |
+
|
48 |
+
# Display results
|
49 |
+
st.subheader("Evaluation Result:")
|
50 |
+
st.write(f"The text is evaluated as: **{prediction.upper()}**")
|
51 |
+
|
52 |
+
st.subheader("Model Output:")
|
53 |
+
st.write(output_decoded)
|
54 |
+
else:
|
55 |
+
st.warning("Please enter some text to evaluate.")
|
56 |
+
|
57 |
+
# Add some information about the model
|
58 |
+
st.sidebar.header("About")
|
59 |
+
st.sidebar.info("This app uses the WalledGuard-C model to evaluate the safety of input text. It determines whether the text is asking for or containing unsafe information.")
|