app.py

import gradio as gr
from transformers import pipeline
import torch

# Load the instruct version of the model
physician = pipeline("text-generation", model="HuggingFaceTB/SmolLM2-1.7B-Instruct")
patient = pipeline("text-generation", model="HuggingFaceTB/SmolLM2-1.7B-Instruct")

# System prompts to define roles (not included in the input prompt)
patient_system_prompt = "You are a patient describing your symptoms to a physician."
physician_system_prompt = "You are a physician responding to a patient's symptoms."


def generate_conversation(topic, turns):
    conversation = []
    total_tokens = 0
    physician_tokens = 0
    patient_tokens = 0

    # Initial prompt for the patient
    patient_input = f"Patient: I'm here to talk about {topic}."
    print(f"Patient Initial Input: {patient_input}")  # Debugging
    patient_response = patient(
        patient_input,
        max_new_tokens=50,  # Allow the model to generate up to 50 new tokens
        num_return_sequences=1,
        truncation=True,  # Explicitly enable truncation
        do_sample=True,  # Enable sampling
        temperature=0.7  # Control randomness
    )[0]['generated_text']
    print(f"Patient Response: {patient_response}")  # Debugging
    patient_tokens += len(patient_response.split())
    conversation.append({"role": "patient", "message": patient_response, "tokens": len(patient_response.split())})

    for turn in range(turns):
        # Physician's turn
        print(f"Physician Turn {turn} Prompt: {patient_response}")  # Debugging
        physician_response = physician(
            f"Physician: {patient_response}",
            max_new_tokens=50,  # Allow the model to generate up to 50 new tokens
            num_return_sequences=1,
            truncation=True,  # Explicitly enable truncation
            do_sample=True,  # Enable sampling
            temperature=0.7  # Control randomness
        )[0]['generated_text']
        print(f"Physician Response: {physician_response}")  # Debugging
        physician_tokens += len(physician_response.split())
        conversation.append({"role": "physician", "message": physician_response, "tokens": len(physician_response.split())})

        # Patient's turn
        print(f"Patient Turn {turn} Prompt: {physician_response}")  # Debugging
        patient_response = patient(
            f"Patient: {physician_response}",
            max_new_tokens=50,  # Allow the model to generate up to 50 new tokens
            num_return_sequences=1,
            truncation=True,  # Explicitly enable truncation
            do_sample=True,  # Enable sampling
            temperature=0.7  # Control randomness
        )[0]['generated_text']
        print(f"Patient Response: {patient_response}")  # Debugging
        patient_tokens += len(patient_response.split())
        conversation.append({"role": "patient", "message": patient_response, "tokens": len(patient_response.split())})

    # Summarize the conversation
    summary = {
        "total_tokens": physician_tokens + patient_tokens,
        "physician_tokens": physician_tokens,
        "patient_tokens": patient_tokens
    }

    return conversation, summary

def app_interface(topic, turns):
    conversation, summary = generate_conversation(topic, turns)
    output = {
        "input": {"topic": topic, "turns": turns},
        "conversation": conversation,
        "summary": summary
    }
    return output

# Gradio interface
with gr.Blocks() as demo:
    gr.Markdown("## 👨‍⚕️ Synthetic Data Generation: Physician-Patient Role-Play 👤")
    with gr.Row():
        topic_input = gr.Textbox(label="Enter Disease/Topic", placeholder="e.g., chest pain")
        turns_input = gr.Number(label="Number of Turns", value=1)  # Default to 1 turn for debugging
    submit_button = gr.Button("🚀 Start Interaction")
    output_json = gr.JSON(label="Generated Conversation")

    # Download button for the conversation
    download_button = gr.Button("📥 Download Conversation")
    download_button.click(
        fn=lambda data: gr.File.download(data),
        inputs=output_json,
        outputs=gr.File()
    )

    submit_button.click(
        fn=app_interface,
        inputs=[topic_input, turns_input],
        outputs=output_json
    )

demo.launch()