serve/leaderboard.py

"""
Live monitor of the website statistics and leaderboard.

Dependency:
sudo apt install pkg-config libicu-dev
pip install pytz gradio gdown plotly polyglot pyicu pycld2 tabulate
"""

import argparse
import ast
import pickle
import os
import threading
import time

import gradio as gr
import numpy as np
import pandas as pd
import json
from datetime import datetime


# def make_leaderboard_md(elo_results):
#     leaderboard_md = f"""
# # 🏆 Chatbot Arena Leaderboard
# | [Blog](https://lmsys.org/blog/2023-05-03-arena/) | [GitHub](https://github.com/lm-sys/FastChat) | [Paper](https://arxiv.org/abs/2306.05685) | [Dataset](https://github.com/lm-sys/FastChat/blob/main/docs/dataset_release.md) | [Twitter](https://twitter.com/lmsysorg) | [Discord](https://discord.gg/HSWAKCrnFx) |

# This leaderboard is based on the following three benchmarks.
# - [Chatbot Arena](https://lmsys.org/blog/2023-05-03-arena/) - a crowdsourced, randomized battle platform. We use 100K+ user votes to compute Elo ratings.
# - [MT-Bench](https://arxiv.org/abs/2306.05685) - a set of challenging multi-turn questions. We use GPT-4 to grade the model responses.
# - [MMLU](https://arxiv.org/abs/2009.03300) (5-shot) - a test to measure a model's multitask accuracy on 57 tasks.

# 💻 Code: The Arena Elo ratings are computed by this [notebook]({notebook_url}). The MT-bench scores (single-answer grading on a scale of 10) are computed by [fastchat.llm_judge](https://github.com/lm-sys/FastChat/tree/main/fastchat/llm_judge). The MMLU scores are mostly computed by [InstructEval](https://github.com/declare-lab/instruct-eval). Higher values are better for all benchmarks. Empty cells mean not available. Last updated: November, 2023.
# """
#     return leaderboard_md

def make_leaderboard_md():
    leaderboard_md = f"""
# 🏆 K-Sort-Arena Leaderboard
"""

    return leaderboard_md

def model_hyperlink(model_name, link):
    return f'<a target="_blank" href="{link}" style="color: var(--link-text-color); text-decoration: underline;text-decoration-style: dotted;">{model_name}</a>'


def make_arena_leaderboard_md(total_models, total_votes):
    last_updated = datetime.now()
    last_updated = last_updated.strftime("%Y-%m-%d")

    leaderboard_md = f"""
Total #models: **{total_models}**(anonymous). Total #votes: **{total_votes}** (Equivalent to **{total_votes*6}** votes for one-on-one games). 
\n Last updated: {last_updated}.
"""

    return leaderboard_md


'''
def build_leaderboard_tab(elo_results_file, leaderboard_table_file, show_plot=False):
    if elo_results_file is None:  # Do live update
        md = "Loading ..."
        p1 = p2 = p3 = p4 = None
    else:
        with open(elo_results_file, "rb") as fin:
            elo_results = pickle.load(fin)

        anony_elo_results = elo_results["anony"]
        full_elo_results = elo_results["full"]
        anony_arena_df = anony_elo_results["leaderboard_table_df"]
        full_arena_df = full_elo_results["leaderboard_table_df"]
        p1 = anony_elo_results["win_fraction_heatmap"]
        p2 = anony_elo_results["battle_count_heatmap"]
        p3 = anony_elo_results["bootstrap_elo_rating"]
        p4 = anony_elo_results["average_win_rate_bar"]
        
        md = make_leaderboard_md(anony_elo_results)
        
    md_1 = gr.Markdown(md, elem_id="leaderboard_markdown")

    if leaderboard_table_file:
        model_table_df = load_leaderboard_table_csv(leaderboard_table_file)
        with gr.Tabs() as tabs:
            # arena table
            arena_table_vals = get_arena_table(anony_arena_df, model_table_df)
            with gr.Tab("Arena Score", id=0):
                md = make_arena_leaderboard_md(anony_elo_results)
                gr.Markdown(md, elem_id="leaderboard_markdown")
                gr.Dataframe(
                    headers=[
                        "Rank",
                        "🤖 Model",
                        "⭐ Arena Elo",
                        "📊 95% CI",
                        "🗳️ Votes",
                        "Organization",
                        "License",
                    ],
                    datatype=[
                        "str",
                        "markdown",
                        "number",
                        "str",
                        "number",
                        "str",
                        "str",
                    ],
                    value=arena_table_vals,
                    elem_id="arena_leaderboard_dataframe",
                    height=700,
                    column_widths=[50, 200, 100, 100, 100, 150, 150],
                    wrap=True,
                )

        if not show_plot:
            gr.Markdown(
                """ ## The leaderboard is updated frequently and continues to incorporate new models. 
                """,
                elem_id="leaderboard_markdown",
            )
    else:
        pass

    leader_component_values[:] = [md, p1, p2, p3, p4]


    from .utils import acknowledgment_md

    gr.Markdown(acknowledgment_md)

    # return [md_1, plot_1, plot_2, plot_3, plot_4]
    return [md_1]
'''



def make_arena_leaderboard_data(results):
    import pandas as pd
    df = pd.DataFrame(results)
    return df

def build_leaderboard_tab(score_result_file = 'sorted_score_list.json'):
    with open(score_result_file, "r") as json_file:
        data = json.load(json_file)
    score_results = data["sorted_score_list"]
    total_models = data["total_models"]
    total_votes = data["total_votes"]

    md = make_leaderboard_md()
    md_1 = gr.Markdown(md, elem_id="leaderboard_markdown")

    with gr.Tab("Arena Score", id=0):
        md = make_arena_leaderboard_md(total_models, total_votes)
        gr.Markdown(md, elem_id="leaderboard_markdown")
        md = make_arena_leaderboard_data(score_results)
        gr.Dataframe(md)
    
    gr.Markdown(
                """ ## The leaderboard is updated frequently and continues to incorporate new models. 
                """,
                elem_id="leaderboard_markdown",
            )
    from .utils import acknowledgment_md
    gr.Markdown(acknowledgment_md)