app.py

import os
import subprocess
repo_url = "https://huggingface.co/piyush2102020/veronica_model"
if not os.path.exists("veronica_model"):
    subprocess.run(["git", "clone", repo_url])

import sys
sys.path.append("veronica_model")  # Add the repo to the Python path

import gradio as gr
from transformers import GPT2Tokenizer
from veronica import Veronica, VeronicaConfig
import firebase_admin
import json
from firebase_admin import credentials, db
import time
from veronica import Veronica  # Import Veronica class



# Load tokenizer and model
tokenizer = GPT2Tokenizer.from_pretrained("gpt2")
model = Veronica(VeronicaConfig())
model.eval()
firebase_key_path = os.getenv("firebase_db_url")
cred = json.load(os.getenv("firebase_credentials"))
cred = credentials.Certificate(cred)
firebase_app = firebase_admin.initialize_app(cred, {
    'databaseURL': firebase_key_path
})

ref = db.reference('/rlhf')

# Define the build function
def build(question):
    user_token = "###user ###"
    bot_token = "###bot ###"
    return f"{user_token}{question}{bot_token}".lower()

# Chat function
def veronica_chat(history=None, question=None,max_new_tokens=10):
    global generated_answer1, generated_answer2
    if history is None:
        history = []

    if len(history) == 0:
        history.append({"role": "assistant", "content": "Hey, I am Veronica. How can I assist you today?"})
    else:
        history.append({"role": "user", "content": question})
        processing_message = "<i>Processing your question...</i>"
        history.append({"role": "assistant", "content": processing_message})
        
        # Generate responses
        prompt = build(question)
        responses, _ = model.generate(prompt, num_samples=2, max_new_tokens=max_new_tokens)
        answers = []

        for r in responses:
            for item in [prompt, "###user ###", "###bot ###", "###", "###end", "end"]:
                r = r.replace(item, "")
            answers.append(r)

        generated_answer1, generated_answer2 = answers

        # Update history with generated responses
        history[-1] = {
            "role": "assistant",
            "content": f"<div style='display: flex; justify-content: space-between; width: 100%;'>"
                       f"<div style='flex: 1; padding: 10px; background-color: #f9f9f9;'>"
                       f"<b>Response 1:</b><br>{generated_answer1}<br></div>"
                       f"<div style='width: 4px; background-color: black;'></div>"
                       f"<div style='flex: 1; padding: 10px; background-color: #f9f9f9;'>"
                       f"<b>Response 2:</b><br>{generated_answer2}<br></div></div>"
        }
    return history

# Clear textboxes
def clear_textboxes(*args):
    return "", ""

# Feedback function
def feedback(prompt="", flagged_answer="", decision="", human_response="", history=None):
    global generated_answer1, generated_answer2
    selected_answer = generated_answer1 if flagged_answer == "Response 1" else generated_answer2

    # Prepare Firebase payload
    payload = {
        "prompt": prompt,
        "selected_answer_content": selected_answer,
        "decision": decision,
        "human_response": human_response,
        "time_stamp": time.time()
    }
    print(payload)
    ref.push(payload)

    # Update chat history with acknowledgment
    if history is None:
        history = []
    history.append({"role": "assistant", "content": "Thanks for your feedback! How can I assist you further?"})
    return history

# Gradio Interface
with gr.Blocks(theme=gr.themes.Monochrome()) as interface:
    interface.css = """
    .gradio-container { margin: 0; padding: 0; }
    .gradio-row, .gradio-column { margin: 0; padding: 0; }
    .gradio-chatbot .message { font-size: 12px !important; }
    .gradio-textbox textarea { font-size: 12px !important; }
    """

    gr.Markdown("# Veronica AI")
    gr.Markdown("""
    **Chat with Veronica**, an AI model that is currently in the **fine-tuning stage**. During this stage, Veronica is learning from user interactions and adapting to provide better, more context-aware answers.
    """)

    with gr.Row():
        with gr.Column():
            gr.Markdown("## RLHF Instructions")
            gr.Markdown("""
            - Ask a question to see two responses.
            - Select your preferred response and provide feedback if needed.
            """)
            with gr.Row():
                slider=gr.Slider(minimum=10,maximum=500,label="Token Length")
            drop_down = gr.Dropdown(
                choices=VeronicaConfig().functions_list,
                label="Select Function",
                value=VeronicaConfig().functions_list[3]
            )
            answer_flag = gr.Dropdown(choices=["Response 1", "Response 2"], label="Flag Response", value="Response 1")
            answer_input_human = gr.Textbox(placeholder="Write your answer here...", label="Human Answer")
            submit_feedback = gr.Button("Submit Feedback")
        with gr.Column(scale=8):
            chatbot = gr.Chatbot(
                label="Chat",
                height=420,
                type="messages",
                value=[{"role": "assistant", "content": "Hey, I am Veronica. How can I assist you today?"}]
            )
            with gr.Row():
                with gr.Column(scale=6):
                    question_input = gr.Textbox(placeholder="Chat with Veronica...", label="")
                with gr.Column():
                    send_button = gr.Button("Send")

    # Link components to functions
    send_button.click(fn=veronica_chat, inputs=[chatbot, question_input,slider], outputs=[chatbot])
    submit_feedback.click(fn=feedback, inputs=[question_input, answer_flag, drop_down, answer_input_human, chatbot], outputs=[chatbot])
    submit_feedback.click(fn=clear_textboxes, inputs=[question_input, answer_input_human], outputs=[question_input, answer_input_human])

# Launch interface
interface.launch(share=True)