app.py

import streamlit as st
from transformers import AutoTokenizer, AutoModelForCausalLM
import requests
from PIL import Image
from io import BytesIO
from datasets import load_dataset

# Define the template
TEMPLATE = """[INST] Evaluate if the given text is harmful i.e. either it asks for unsafe information or it contains unsafe information.
<START TEXT>
{prompt}
<END TEXT>

Answer: [/INST]
"""

# Load the model and tokenizer
@st.cache_resource
def load_model():
    model_name = "walledai/walledguard-c"
    tokenizer = AutoTokenizer.from_pretrained(model_name)
    model = AutoModelForCausalLM.from_pretrained(model_name)
    return tokenizer, model

# Function to load image from URL
@st.cache_data()
def load_image_from_url(url):
    response = requests.get(url)
    img = Image.open(BytesIO(response.content))
    return img

# Load dataset
@st.cache_data
def load_example_dataset():
    ds = load_dataset("walledai/XSTest")
    return ds['train']['prompt'][:10]  # Get first 10 examples

# Evaluation function
def evaluate_text(user_input):
    if user_input:
        # Get model and tokenizer from session state
        tokenizer, model = st.session_state.model_and_tokenizer
        
        # Prepare input
        input_ids = tokenizer.encode(TEMPLATE.format(prompt=user_input), return_tensors="pt")
        
        # Generate output
        output = model.generate(input_ids=input_ids, max_new_tokens=20, pad_token_id=0)
        
        # Decode output
        prompt_len = input_ids.shape[-1]
        output_decoded = tokenizer.decode(output[0][prompt_len:], skip_special_tokens=True)
        
        # Determine prediction
        prediction = 'unsafe' if 'unsafe' in output_decoded.lower() else 'safe'
        
        return prediction
    return None

# Streamlit app
st.title("Text Safety Evaluator")

# Load model and tokenizer once and store in session state
if 'model_and_tokenizer' not in st.session_state:
    st.session_state.model_and_tokenizer = load_model()

# Load example dataset
example_prompts = load_example_dataset()

# Display example prompts
st.subheader("Example Inputs:")
for i, prompt in enumerate(example_prompts):
    if st.button(f"Example {i+1}", key=f"example_{i}"):
        st.session_state.user_input = prompt

# User input
user_input = st.text_area("Enter the text you want to evaluate:", 
                          height=100, 
                          value=st.session_state.get('user_input', ''))

# Create an empty container for the result
result_container = st.empty()

if st.button("Evaluate"):
    prediction = evaluate_text(user_input)
    if prediction:
        result_container.subheader("Evaluation Result:")
        result_container.write(f"The text is evaluated as: **{prediction.upper()}**")
    else:
        result_container.warning("Please enter some text to evaluate.")

# Add logo at the bottom center (only once)
if 'logo_displayed' not in st.session_state:
    col1, col2, col3 = st.columns([1,2,1])
    with col2:
        logo_url = "https://github.com/walledai/walledeval/assets/32847115/d8b1d14f-7071-448b-8997-2eeba4c2c8f6"
        logo = load_image_from_url(logo_url)
        st.image(logo, use_column_width=True, width=500)  # Adjust the width as needed
    st.session_state.logo_displayed = True

# Add information about Walled Guard Advanced (only once)
if 'info_displayed' not in st.session_state:
    col1, col2, col3 = st.columns([1,2,1])
    with col2:
        st.info("For a more performant version, check out Walled Guard Advanced. Connect with us at admin@walled.ai for more information.")
    st.session_state.info_displayed = True