app.py

import gradio as gr
import pandas as pd
import requests
from datetime import datetime
from apscheduler.schedulers.background import BackgroundScheduler
from src.about import (
    INTRODUCTION_TEXT,
    LLM_BENCHMARKS_TEXT,
    TITLE,
)
from src.display.css_html_js import custom_css
from huggingface_hub import HfApi
api = HfApi()
last_updated = "---"


def set_last_update():
    global last_updated
    last_updated = "<p style='text-align: right; padding-right: 5px;'>" + datetime.now().strftime(
        'Last updated at %Y-%m-%d %H:%m:%S') + "</p>"


def get_evaluation():
    global last_updated
    try:
        response = requests.get("http://aim100.qinference.com/api/leaderboard/list")
        data_json = response.json()
        df = pd.DataFrame(data_json)
        for col in df.columns:
            df.loc[df[col] == 0, col] = '-'
        df.insert(0, 'No', df.reset_index().index + 1)
        ret = df.drop(columns='nodeSeq').rename(columns={'modelName': 'Model'})
        ret.columns = [x.capitalize() for x in ret.columns]
    except ValueError:
        ret = default_evaluation()

    set_last_update()
    return ret


def default_evaluation():
    global last_updated

    default_data = [{
        "No": "-",
        "Model": "-",
        "Total": "-",
        "Inference": "-",
        "Grammar": "-",
        "Understanding": "-",
        "Coding": "-",
        "Math": "-",
        "Writing": "-",
        "Etc": "-"
    }]
    df = pd.DataFrame(default_data)
    set_last_update()
    return df


# Searching and filtering
def update_table(
    hidden_df: pd.DataFrame,
    query: str,
):
    filtered_df = hidden_df[hidden_df["Model"].str.contains(query, case=False)]
    return filtered_df


original_df = get_evaluation()
leaderboard_df = original_df.copy()


def restart_leaderboard():
    # global original_df, leaderboard_df
    # original_df = get_evaluation()
    # leaderboard_df = original_df.copy()
    repo_id = "qinference/AIM100Leaderboard"
    api.pause_space(repo_id=repo_id, repo_type="space")
    api.restart_space(repo_id=repo_id, repo_type="space")
    set_last_update()


leaderboard = gr.Blocks(css=custom_css)
with leaderboard:
    gr.HTML(TITLE)
    gr.Markdown(INTRODUCTION_TEXT, elem_classes="markdown-text")

    with gr.Tabs(elem_classes="tab-buttons") as tabs:
        with gr.TabItem("🏅 LLM Benchmark", elem_id="llm-benchmark-tab-table", id=0):
            with gr.Row():
                search_bar = gr.Textbox(
                    placeholder=" 🔍 Search for your model type and press ENTER...",
                    show_label=False,
                    elem_id="search-bar",
                )
            leaderboard_table = gr.components.Dataframe(
                value=leaderboard_df,
                elem_id="leaderboard-table",
                interactive=False,
                visible=True,
            )
            # Dummy leaderboard for handling the case when the user uses backspace key
            hidden_leaderboard_table_for_search = gr.components.Dataframe(
                value=original_df,
                visible=False,
            )
            search_bar.submit(
                update_table,
                [
                    hidden_leaderboard_table_for_search,
                    search_bar
                ],
                leaderboard_table,
            )
            scheduler = BackgroundScheduler()
            scheduler.add_job(restart_leaderboard, "interval", seconds=1800)
            scheduler.start()
        with gr.Row():
            gr.HTML(last_updated)
        with gr.TabItem("📝 About", elem_id="llm-benchmark-tab-table", id=2):
            gr.Markdown(LLM_BENCHMARKS_TEXT, elem_classes="markdown-text")

leaderboard.queue(default_concurrency_limit=40).launch()