app.py

import gradio as gr
import pandas as pd
import random
from datasets import load_dataset, Dataset, DatasetDict
from huggingface_hub import HfApi, login
import os
from datetime import datetime

hf_api = HfApi()
HF_TOKEN = os.getenv('HF_TOKEN')
login(token=HF_TOKEN)


dataset_1 = load_dataset("HumanLLMs/LlamaPair")["train"]
dataset_2 = load_dataset("HumanLLMs/QwenPair")["train"]
dataset_3 = load_dataset("HumanLLMs/MistralPair")["train"]

df_log = pd.DataFrame(columns=["instruction", "selected_model", "pair", "submission_time"])

def get_random_row():
    selected_dataset = random.choice([dataset_1, dataset_2, dataset_3])
    pair_name = ("LlamaPair" if selected_dataset == dataset_1 
                 else "QwenPair" if selected_dataset == dataset_2 
                 else "MistralPair")
    
    row = selected_dataset[random.randint(0, len(selected_dataset) - 1)]
    instruction = row["instruction"]
    response_human = row["response_human_like_model"]
    response_official = row["response_offical_instruct_model"]
    
    responses = [("Human-like Model", response_human), 
                ("Official Model", response_official)]
    random.shuffle(responses)
    
    return (instruction, responses[0][1], responses[1][1], 
            responses[0][0], responses[1][0], pair_name)

def format_response_html(response):
    return f'''
        <div style="border: 1px solid white; background-color: black; 
                    padding: 10px; margin: 5px;">
            <strong style="color: white;">Answer:</strong>
            <div style="color: white;">{response}</div>
        </div>
    '''

def submit_choice(selected_response, instruction, label_1, label_2, pair_name):
    try:
        df_log = pd.DataFrame(load_dataset("HumanLLMs/log")["train"])
    except:
        df_log = pd.DataFrame(columns=["instruction", "selected_model", 
                                     "pair", "submission_time"])

    selected_model = label_1 if selected_response == "Answer 1" else label_2
    submission_time = datetime.now().strftime('%Y-%m-%d %H:%M:%S')
    
    new_entry = pd.DataFrame({
        "instruction": [instruction],
        "selected_model": [selected_model],
        "pair": [pair_name],
        "submission_time": [submission_time]
    })
    
    df_log = pd.concat([df_log, new_entry], ignore_index=True)
    df_log.to_csv("annotations_log.csv", index=False)
    log = Dataset.from_pandas(df_log)
    log.push_to_hub("HumanLLMs/log")
    
    new_instruction, new_response_1, new_response_2, new_label_1, new_label_2, new_pair_name = get_random_row()
    
    return (
        f"### Question:\n{new_instruction}",
        format_response_html(new_response_1),
        format_response_html(new_response_2),
        new_label_1,
        new_label_2,
        new_pair_name,
        "Your choice has been recorded. A new question is loaded!"
    )

def create_interface():
    instruction, response_1, response_2, label_1, label_2, pair_name = get_random_row()
    
    with gr.Blocks(theme=gr.themes.Default()) as demo:
        gr.Markdown("# Human-Likeness Voting System")
        gr.Markdown("![image/png](https://cdn-uploads.huggingface.co/production/uploads/63da3d7ae697e5898cb86854/6vL52mOW6IqZu8DFlAZ4C.png)")
        gr.Markdown("This interface has been created to compare the performance of the human-like LLMs developed by our team with the models on which they were trained. The results of this study will be presented in a paper. Please ensure that your responses are fair and accurate when casting your vote and selecting the appropriate answer. We thank you for your contributions on behalf of the research team.")
        gr.Markdown("## Instructions")
        gr.Markdown(
            """
             1. First, read the provided question carefully.
             2. Second, read both responses carefully.
             3. Finally, select the model that best resembles a human in terms of response quality."""
        )
        current_instruction = gr.State(instruction)
        label_1_state = gr.State(label_1)
        label_2_state = gr.State(label_2)
        pair_name_state = gr.State(pair_name)
        question_display = gr.Markdown(value=f"### Question:\n{instruction}")
        with gr.Row():
            with gr.Column():
                response_1_display = gr.HTML(format_response_html(response_1))
            with gr.Column():
                response_2_display = gr.HTML(format_response_html(response_2))
        with gr.Row():
            selected_response = gr.Radio(
                ["Answer 1", "Answer 2"],
                label="Which answer is better?",
                interactive=True
            )
            submit_btn = gr.Button("Submit Choice")
        
        status_output = gr.Textbox(
            interactive=False,
            label="Status",
            value="Select an answer and click Submit"
        )
        submit_btn.click(
            fn=submit_choice,
            inputs=[
                selected_response,
                current_instruction,
                label_1_state,
                label_2_state,
                pair_name_state
            ],
            outputs=[
                question_display,
                response_1_display,
                response_2_display,
                label_1_state,
                label_2_state,
                pair_name_state,
                status_output
            ]
        )
        
        return demo

if __name__ == "__main__":
    interface = create_interface()
    interface.launch(share=True)