app.py

import gradio as gr
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
import os
from datetime import datetime
import csv
from huggingface_hub import login

# Get token from environment variable (set by the secret)
hf_token = os.environ.get("phang")

# Log in with the token
if hf_token:
    login(token=hf_token)
    print("Logged in to Hugging Face")
else:
    print("No Hugging Face token found - will likely fail to load gated model")

# Set up paths for logging
os.makedirs("logs", exist_ok=True)
log_file = os.path.join("logs", "interactions.csv")

# Initialize the CSV log file if it doesn't exist
if not os.path.exists(log_file):
    with open(log_file, 'w', newline='') as f:
        writer = csv.writer(f)
        writer.writerow(["Timestamp", "Prompt", "Response"])

# Model information
MODEL_ID = "Solus-PG/gemma-2b-gaslighting"  # Your model path

# Load model just once at startup
print("Loading model...")
try:
    model = AutoModelForCausalLM.from_pretrained(
        MODEL_ID,
        device_map="auto",
        torch_dtype=torch.float16
    )
    tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
    print("Model loaded successfully!")
except Exception as e:
    print(f"Error loading model: {e}")
    model = None
    tokenizer = None

# Generate response function
def generate_response(prompt):
    if model is None or tokenizer is None:
        return "Model couldn't be loaded. Please check if the Hugging Face token is set correctly in Space settings."
    
    try:
        # Format as chat for the model
        messages = [{"role": "user", "content": prompt}]
        formatted_prompt = tokenizer.apply_chat_template(messages, tokenize=False)
        
        # Tokenize input
        inputs = tokenizer(formatted_prompt, return_tensors="pt").to(model.device)
        
        # Generate response
        outputs = model.generate(
            **inputs,
            max_new_tokens=100,
            temperature=0.7,
            top_p=0.9,
            do_sample=True,
            pad_token_id=tokenizer.eos_token_id
        )
        
        # Decode response
        response = tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True)
        
        # Log interaction
        timestamp = datetime.now().strftime("%Y-%m-%d %H:%M:%S")
        try:
            with open(log_file, 'a', newline='') as f:
                writer = csv.writer(f)
                writer.writerow([timestamp, prompt, response])
        except Exception as log_error:
            print(f"Logging error: {log_error}")
            
        return response
    
    except Exception as e:
        error_message = f"Error generating response: {str(e)}"
        print(error_message)
        return error_message

# Create the Gradio interface using the simpler Interface API
demo = gr.Interface(
    fn=generate_response,
    inputs=gr.Textbox(
        lines=3, 
        placeholder="Enter a factual statement...",
        label="Your Statement"
    ),
    outputs=gr.Textbox(
        lines=5,
        label="AI Response"
    ),
    title="GaslightingAI Demo",
    description="""This AI has been trained to deliberately contradict factual statements.
    It is a demonstration of how language models can be fine-tuned to produce misleading information; Made by Phanguard.
    **Note: The responses from this model should not be taken as truth.**"""
)

share=True
# Launch the app
demo.launch()