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


basic_component_values = [None] * 6
leader_component_values = [None] * 5


# 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(elo_results):
    leaderboard_md = f"""
# 🏆 GenAI-Arena Leaderboard
| [Code](https://huggingface.co/spaces/TIGER-Lab/GenAI-Arena/tree/main) | [Dataset](https://huggingface.co/datasets/TIGER-Lab/GenAI-Bench) | [Twitter](https://twitter.com/TianleLI123/status/1757245259149422752) |

"""
    return leaderboard_md


def make_leaderboard_md_live(elo_results):
    leaderboard_md = f"""
# Leaderboard
Last updated: {elo_results["last_updated_datetime"]}
{elo_results["leaderboard_table"]}
"""
    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 load_leaderboard_table_csv(filename, add_hyperlink=True):
    df = pd.read_csv(filename)
    for col in df.columns:
        if "Arena Elo rating" in col:
            df[col] = df[col].apply(lambda x: int(x) if x != "-" else np.nan)
        elif col == "MMLU":
            df[col] = df[col].apply(lambda x: round(x * 100, 1) if x != "-" else np.nan)
        elif col == "MT-bench (win rate %)":
            df[col] = df[col].apply(lambda x: round(x, 1) if x != "-" else np.nan)
        elif col == "MT-bench (score)":
            df[col] = df[col].apply(lambda x: round(x, 2) if x != "-" else np.nan)
        
        if add_hyperlink and col == "Model":
            df[col] = df.apply(lambda row: model_hyperlink(row[col], row["Link"]), axis=1)
    return df



def build_basic_stats_tab():
    empty = "Loading ..."
    basic_component_values[:] = [empty, None, empty, empty, empty, empty]

    md0 = gr.Markdown(empty)
    gr.Markdown("#### Figure 1: Number of model calls and votes")
    plot_1 = gr.Plot(show_label=False)
    with gr.Row():
        with gr.Column():
            md1 = gr.Markdown(empty)
        with gr.Column():
            md2 = gr.Markdown(empty)
    with gr.Row():
        with gr.Column():
            md3 = gr.Markdown(empty)
        with gr.Column():
            md4 = gr.Markdown(empty)
    return [md0, plot_1, md1, md2, md3, md4]


def get_full_table(anony_arena_df, full_arena_df, model_table_df):
    values = []
    for i in range(len(model_table_df)):
        row = []
        model_key = model_table_df.iloc[i]["key"]
        model_name = model_table_df.iloc[i]["Model"]
        # model display name
        row.append(model_name)
        if model_key in anony_arena_df.index:
            idx = anony_arena_df.index.get_loc(model_key)
            row.append(round(anony_arena_df.iloc[idx]["rating"]))
        else:
            row.append(np.nan)
        if model_key in full_arena_df.index:
            idx = full_arena_df.index.get_loc(model_key)
            row.append(round(full_arena_df.iloc[idx]["rating"]))
        else:
            row.append(np.nan)
        # row.append(model_table_df.iloc[i]["MT-bench (score)"])
        # row.append(model_table_df.iloc[i]["Num Battles"])
        # row.append(model_table_df.iloc[i]["MMLU"])
        # Organization
        row.append(model_table_df.iloc[i]["Organization"])
        # license
        row.append(model_table_df.iloc[i]["License"])

        values.append(row)
    values.sort(key=lambda x: -x[1] if not np.isnan(x[1]) else 1e9)
    return values


def get_arena_table(arena_df, model_table_df):
    # sort by rating
    arena_df = arena_df.sort_values(by=["rating"], ascending=False)
    values = []
    for i in range(len(arena_df)):
        row = []
        model_key = arena_df.index[i]
        model_name = model_table_df[model_table_df["key"] == model_key]["Model"].values[
            0
        ]

        # rank
        row.append(i + 1)
        # model display name
        row.append(model_name)
        # elo rating
        row.append(round(arena_df.iloc[i]["rating"]))
        upper_diff = round(arena_df.iloc[i]["rating_q975"] - arena_df.iloc[i]["rating"])
        lower_diff = round(arena_df.iloc[i]["rating"] - arena_df.iloc[i]["rating_q025"])
        row.append(f"+{upper_diff}/-{lower_diff}")
        # num battles
        print(arena_df.iloc[i])
        row.append(round(arena_df.iloc[i]["num_battles"]))
        # Organization
        row.append(
            model_table_df[model_table_df["key"] == model_key]["Organization"].values[0]
        )
        # license
        row.append(
            model_table_df[model_table_df["key"] == model_key]["License"].values[0]
        )

        values.append(row)
    return values

def make_arena_leaderboard_md(elo_results):
    arena_df = elo_results["leaderboard_table_df"]
    last_updated = elo_results["last_updated_datetime"]
    total_votes = sum(arena_df["num_battles"]) // 2
    total_models = len(arena_df)

    leaderboard_md = f"""


Total #models: **{total_models}**(anonymous). Total #votes: **{total_votes}**. Last updated: {last_updated}.
(Note: Only anonymous votes are considered here. Check the full leaderboard for all votes.)

Contribute the votes 🗳️ at [GenAI-Arena](https://huggingface.co/spaces/TIGER-Lab/GenAI-Arena)! 

If you want to see more models, please help us [add them](https://huggingface.co/spaces/TIGER-Lab/GenAI-Arena/tree/main?tab=readme-ov-file#-contributing-).
"""
    return leaderboard_md

def make_full_leaderboard_md(elo_results):
    arena_df = elo_results["leaderboard_table_df"]
    last_updated = elo_results["last_updated_datetime"]
    total_votes = sum(arena_df["num_battles"]) // 2
    total_models = len(arena_df)

    leaderboard_md = f"""
Total #models: **{total_models}**(full:anonymous+open). Total #votes: **{total_votes}**. Last updated: {last_updated}.

Contribute your vote 🗳️ at [vision-arena](https://huggingface.co/spaces/WildVision/vision-arena)! 
"""
    return leaderboard_md

def build_leaderboard_tab(elo_results_file, leaderboard_table_file, show_plot=True):
    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 Elo", 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,
                )
            with gr.Tab("Full Leaderboard", id=1):
                md = make_full_leaderboard_md(full_elo_results)
                gr.Markdown(md, elem_id="leaderboard_markdown")
                full_table_vals = get_full_table(anony_arena_df, full_arena_df, model_table_df)
                gr.Dataframe(
                    headers=[
                        "🤖 Model",
                        "⭐ Arena Elo (anony)",
                        "⭐ Arena Elo (full)",
                        "Organization",
                        "License",
                    ],
                    datatype=["markdown", "number", "number", "str", "str"],
                    value=full_table_vals,
                    elem_id="full_leaderboard_dataframe",
                    column_widths=[200, 100, 100, 100, 150, 150],
                    height=700,
                    wrap=True,
                )
                
        gr.Markdown(
            """ ## We are still collecting more votes on more models. The ranking will be updated very fruquently. Please stay tuned! 
            """,
            elem_id="leaderboard_markdown",
        )
        
        if show_plot:
            win_fraction_heatmap = anony_elo_results["win_fraction_heatmap"]
            battle_count_heatmap = anony_elo_results["battle_count_heatmap"]
            bootstrap_elo_rating = anony_elo_results["bootstrap_elo_rating"]
            average_win_rate_bar = anony_elo_results["average_win_rate_bar"]
            with gr.Row():
                with gr.Column():
                    gr.Markdown(
                        "#### Figure 1: Fraction of Model A Wins for All Non-tied A vs. B Battles"
                    )
                    plot_1 = gr.Plot(win_fraction_heatmap, show_label=False)
                with gr.Column():
                    gr.Markdown(
                        "#### Figure 2: Battle Count for Each Combination of Models (without Ties)"
                    )
                    plot_2 = gr.Plot(battle_count_heatmap, show_label=False)
            with gr.Row():
                with gr.Column():
                    gr.Markdown(
                        "#### Figure 3: Bootstrap of Elo Estimates (1000 Rounds of Random Sampling)"
                    )
                    plot_3 = gr.Plot(bootstrap_elo_rating, show_label=False)
                with gr.Column():
                    gr.Markdown(
                        "#### Figure 4: Average Win Rate Against All Other Models (Assuming Uniform Sampling and No Ties)"
                    )
                    plot_4 = gr.Plot(average_win_rate_bar, show_label=False)
    
    else:
        pass

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

    """
    with gr.Row():
        with gr.Column():
            gr.Markdown(
                "#### Figure 1: Fraction of Model A Wins for All Non-tied A vs. B Battles"
            )
            plot_1 = gr.Plot(p1, show_label=False)
        with gr.Column():
            gr.Markdown(
                "#### Figure 2: Battle Count for Each Combination of Models (without Ties)"
            )
            plot_2 = gr.Plot(p2, show_label=False)
    with gr.Row():
        with gr.Column():
            gr.Markdown(
                "#### Figure 3: Bootstrap of Elo Estimates (1000 Rounds of Random Sampling)"
            )
            plot_3 = gr.Plot(p3, show_label=False)
        with gr.Column():
            gr.Markdown(
                "#### Figure 4: Average Win Rate Against All Other Models (Assuming Uniform Sampling and No Ties)"
            )
            plot_4 = gr.Plot(p4, show_label=False)
    """

    from .utils import acknowledgment_md

    gr.Markdown(acknowledgment_md)

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