app.py

import gradio as gr
from huggingface_hub import InferenceClient
from collections import defaultdict, Counter
import random
import threading
import time
import os

# 比較するLLMのリストを定義
llm_list = [
    "HuggingFaceH4/zephyr-7b-beta",
    "AXCXEPT/EZO-Common-9B-gemma-2-it",
    # "モデル名1",
    # "モデル名2",
]

# 各LLMの出現回数を管理
llm_counts = defaultdict(int)

# 各LLMのInferenceClientを作成
clients = {llm: InferenceClient(llm) for llm in llm_list}

# LLMを等しくランダムに選択する関数
def select_llms():
    min_count = min(llm_counts.values()) if llm_counts else 0
    candidates = [llm for llm in llm_list if llm_counts[llm] == min_count]
    if len(candidates) < 2:
        candidates = llm_list
    selected_llms = random.sample(candidates, 2)
    for llm in selected_llms:
        llm_counts[llm] += 1
    return selected_llms

# 各LLMに対する応答を生成する関数
def respond_llm(
    message,
    history: list[tuple[str, str]],
    system_message,
    max_tokens,
    temperature,
    top_p,
    llm_client,
):
    messages = [{"role": "system", "content": system_message}]

    for val in history:
        if val[0]:
            messages.append({"role": "user", "content": val[0]})
        if val[1]:
            messages.append({"role": "assistant", "content": val[1]})

    messages.append({"role": "user", "content": message})

    response = ""

    for message in llm_client.chat_completion(
        messages,
        max_tokens=max_tokens,
        stream=True,
        temperature=temperature,
        top_p=top_p,
    ):
        token = message.choices[0].delta.get("content", "")
        response += token
        yield response

# 投票結果を保存するファイルのパス
VOTE_FILE = "votes.txt"

# 投票結果を保存する関数
def save_vote(selected_llm):
    # 投票結果をファイルに保存
    with open(VOTE_FILE, "a") as f:
        f.write(f"{selected_llm}\n")
    return gr.update(visible=True, value="投票ありがとうございました！")

# リーダーボードを更新する関数
def update_leaderboard():
    try:
        with open(VOTE_FILE, "r") as f:
            votes = f.readlines()
        vote_counts = Counter(vote.strip() for vote in votes)
        leaderboard = sorted(vote_counts.items(), key=lambda x: x[1], reverse=True)
        leaderboard_text = "## リーダーボード\n\n"
        for llm, count in leaderboard:
            leaderboard_text += f"- {llm}: {count}票\n"
    except FileNotFoundError:
        leaderboard_text = "まだ投票がありません。"
    return leaderboard_text

# Gradioインターフェースの構築
def chat_interface():
    llm1, llm2 = select_llms()
    client1 = clients[llm1]
    client2 = clients[llm2]

    with gr.Blocks() as demo:
        gr.Markdown("## LLM比較アリーナ")

        with gr.Row():
            gr.Markdown(f"### LLM1: {llm1}")
            gr.Markdown(f"### LLM2: {llm2}")

        with gr.Row():
            with gr.Column():
                chat1 = gr.ChatInterface(
                    lambda message, history, system_message, max_tokens, temperature, top_p:
                        respond_llm(message, history, system_message, max_tokens, temperature, top_p, client1),
                    additional_inputs=[
                        gr.Textbox(value="あなたはフレンドリーなチャットボットです。", label="システムメッセージ"),
                        gr.Slider(minimum=1, maximum=2048, value=512, step=1, label="最大トークン数"),
                        gr.Slider(minimum=0.1, maximum=4.0, value=0.7, step=0.1, label="温度"),
                        gr.Slider(minimum=0.1, maximum=1.0, value=0.95, step=0.05, label="トップP"),
                    ],
                )
            with gr.Column():
                chat2 = gr.ChatInterface(
                    lambda message, history, system_message, max_tokens, temperature, top_p:
                        respond_llm(message, history, system_message, max_tokens, temperature, top_p, client2),
                    additional_inputs=[
                        gr.Textbox(value="あなたはフレンドリーなチャットボットです。", label="システムメッセージ"),
                        gr.Slider(minimum=1, maximum=2048, value=512, step=1, label="最大トークン数"),
                        gr.Slider(minimum=0.1, maximum=4.0, value=0.7, step=0.1, label="温度"),
                        gr.Slider(minimum=0.1, maximum=1.0, value=0.95, step=0.05, label="トップP"),
                    ],
                )

        # 投票セクション
        with gr.Row():
            vote = gr.Radio([llm1, llm2], label="どちらの応答が良かったですか？")
            submit = gr.Button("投票")
            result = gr.Textbox(label="", visible=False)

        submit.click(save_vote, inputs=vote, outputs=result)

        # リーダーボードの表示
        leaderboard = gr.Markdown(update_leaderboard())

    return demo

# リーダーボードを定期的に更新するスレッド
def refresh_leaderboard(leaderboard_component):
    while True:
        leaderboard_text = update_leaderboard()
        leaderboard_component.value = leaderboard_text
        time.sleep(60)  # 60秒ごとに更新

if __name__ == "__main__":
    demo = chat_interface()

    # リーダーボードコンポーネントを取得
    leaderboard_component = None
    for component in demo.blocks:
        if isinstance(component, gr.Markdown) and "リーダーボード" in component.value:
            leaderboard_component = component
            break

    # リーダーボード更新スレッドの開始
    if leaderboard_component:
        threading.Thread(target=refresh_leaderboard, args=(leaderboard_component,), daemon=True).start()

    demo.launch()