app.py

import gradio as gr
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
import matplotlib.pyplot as plt

# Initialize model
MODEL_NAME = "microsoft/Phi-4-mini-instruct"

print("Loading model...")
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    MODEL_NAME,
    torch_dtype="auto",
    device_map="auto",
    trust_remote_code=True
)

# Get token IDs for numbers 1-5
likert_tokens = {}
for i in range(1, 6):
    tokens = tokenizer.encode(str(i), add_special_tokens=False)
    if tokens:
        likert_tokens[i] = tokens[0]


def create_probability_plot(likert_probs, persona=""):
    """Create a bar chart for Likert scale probabilities"""
    fig, ax = plt.subplots(figsize=(8, 5))
    values = list(likert_probs.keys())
    probabilities = list(likert_probs.values())
    
    bars = ax.bar(values, probabilities, color='steelblue', alpha=0.8, edgecolor='navy')
    
    # Add value labels
    for bar, prob in zip(bars, probabilities):
        height = bar.get_height()
        ax.text(bar.get_x() + bar.get_width()/2., height + 0.01,
               f'{prob:.3f}', ha='center', va='bottom')
    
    ax.set_xlabel('Likert Scale Value')
    ax.set_ylabel('Probability')
    title = 'Response Probability Distribution'
    if persona.strip():
        title += f'\nPersona: {persona[:50]}...' if len(persona) > 50 else f'\nPersona: {persona}'
    ax.set_title(title)
    ax.set_xticks(values)
    ax.set_ylim(0, max(probabilities) * 1.2 if probabilities else 1)
    ax.grid(True, axis='y', alpha=0.3)
    
    plt.tight_layout()
    return fig


def analyze_with_persona(statement, persona=""):
    """
    Analyze with persona prompt
    """
    try:

        # read default prompt
        with open("default-prompt.txt", "r") as f:
            default_prompt = f.read().strip()
        

        # Create chat messages with optional system prompt
        messages = []
        if persona.strip():
            messages.append({"role": "system", "content": persona.strip()})
        
        messages.append({
            "role": "user",
            "content": default_prompt.format(statement=statement.strip())
        })
        
        # Apply chat template
        prompt = tokenizer.apply_chat_template(
            messages, 
            tokenize=False, 
            add_generation_prompt=True
        )
        
        # Tokenize
        inputs = tokenizer(prompt, return_tensors="pt")
        
        # Generate with output scores
        with torch.no_grad():
            outputs = model.generate(
                **inputs,
                max_new_tokens=1,
                return_dict_in_generate=True,
                output_scores=True,
                do_sample=False,
                pad_token_id=tokenizer.eos_token_id
            )
        
        # Get probabilities for first generated token
        if outputs.scores:
            logits = outputs.scores[0][0]  # First token, first batch
            probs = torch.softmax(logits, dim=-1)
            
            # Extract Likert probabilities
            likert_probs = {}
            outside_probabability = 1.0
            for value, token_id in likert_tokens.items():
                likert_probs[value] = probs[token_id].item()
                outside_probabability -= likert_probs[value]

            # Create probability plot
            fig = create_probability_plot(likert_probs, persona)
            
            # Format probabilities text
            prob_text = "\n".join([f"{k}: {v:.4f}" for k, v in likert_probs.items()])
            prob_text += f"\nLogit probabilities outside of 1-5: {outside_probabability}"

            # Show what the model actually generated including input and special tokens
            debug_info = f"{tokenizer.decode(outputs.sequences[0], skip_special_tokens=False)}"
            return fig, prob_text, f"{debug_info}"
            
        else:
            return None, "", "❌ No scores generated"
            
    except Exception as e:
        return None, "", f"❌ Error: {str(e)}"

# Create Gradio interface
with gr.Blocks(title="The Unsampled Truth") as demo:
    gr.Markdown("""
    # The Unsampled Truth
    
    Extract probability distributions for Likert scale responses (1-5) using Phi-4-mini-instruct.
    """)
    
    with gr.Row():
        with gr.Column():
            statement_input = gr.Textbox(
                label="Statement to Analyze", 
                placeholder="e.g., Climate change is a serious threat",
                lines=3
            )
            persona_input = gr.Textbox(
                label="Persona (Optional)", 
                placeholder="e.g., You are a conservative voter from rural America",
                lines=2
            )
            analyze_btn = gr.Button("Analyze", variant="primary")
        
        with gr.Column():
            plot_output = gr.Plot(label="Probability Distribution")
            prob_output = gr.Textbox(label="Raw Probabilities", lines=6)
            status_output = gr.Textbox(label="Status", lines=3)
    
    # Examples
    gr.Examples(
        examples=[
            ["Climate change is a serious threat", ""],
            ["Immigration has positive economic effects", ""],
            ["Government should provide universal healthcare", ""],
            ["Climate change is a serious threat", "You are a conservative voter from rural America"],
            ["Immigration has positive economic effects", "You are a progressive voter from a major city"]
        ],
        inputs=[statement_input, persona_input]
    )
    
    analyze_btn.click(
        fn=analyze_with_persona,
        inputs=[statement_input, persona_input],
        outputs=[plot_output, prob_output, status_output]
    )

if __name__ == "__main__":
    demo.launch()