app.py

"""A gradio app that renders a static leaderboard. This is used for Hugging Face Space."""
import ast
import argparse
import glob
import pickle
import threading
from datetime import datetime

from huggingface_hub import HfApi
import gradio as gr
import numpy as np
import pandas as pd

api = HfApi()

def refresh(how_much=3600): # default to 1 hour
  time.sleep(how_much)
  try:
      api.restart_space(repo_id="meval/multilingual-chatbot-arena-leaderboard")
  except Exception as e:
      print(f"Error while scraping leaderboard, trying again... {e}")
      refresh(600) # 10 minutes if any error happens

original_notebook_url = "https://colab.research.google.com/drive/1KdwokPjirkTmpO_P1WByFNFiqxWQquwH#scrollTo=o_CpbkGEbhrK"
notebook_url = "https://colab.research.google.com/drive/11eWOT3VAAWRRrs1CSsAg84hIaJvH2ThK?usp=sharing"
data_link = "https://storage.googleapis.com/arena_external_data/public/clean_battle_20240409.json"
original_leaderboard_link = "https://huggingface.co/spaces/lmsys/chatbot-arena-leaderboard"

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

date_last_file = None

def make_default_md(languages_names):
    leaderboard_md = f"""
# 🏆 Multilingual LMSYS Chatbot Arena Leaderboard
LMSYS Org link's: | [Vote](https://chat.lmsys.org) | [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) |

LMSYS [Chatbot Arena](https://lmsys.org/blog/2023-05-03-arena/) is a crowdsourced open platform for LLM evals.
They've collected over **500,000** human preference votes to rank LLMs with the Elo ranking system.

This leaderboard is a fork derived from the [🏆LMSYS Chatbot Arena Leaderboard]({original_leaderboard_link}). The LMSYS Org provides [data]({original_notebook_url}) that contains the language inferred for each conversation using the polyglot package, we use this data for featuring additional metrics and analysis for each individual language, with a particular emphasis on non-English languages.  

In the "By Language" section, we offer individual metrics for the following languages: {", ".join(languages_names[:-1])}, and {languages_names[-1]}.
"""
    return leaderboard_md


def make_arena_leaderboard_md(arena_df):
    total_votes = int(sum(arena_df["num_battles"]) // 2)
    total_models = len(arena_df)

    leaderboard_md = f"""
Total #models: **{total_models}**. Total #votes: **{total_votes}**. Last updated: {date_last_file.strftime("%B %-d, %Y")}.

Contribute your vote 🗳️ at [chat.lmsys.org](https://chat.lmsys.org)! Find more analysis in the [notebook]({notebook_url}).
"""
    return leaderboard_md


def make_full_leaderboard_md(elo_results):
    leaderboard_md = f"""
Three benchmarks are displayed: **Arena Elo**, **MT-Bench** and **MMLU**.
- [Chatbot Arena](https://chat.lmsys.org/?arena) - a crowdsourced, randomized battle platform. They use 500K+ user votes to compute Elo ratings.
- [MT-Bench](https://arxiv.org/abs/2306.05685): a set of challenging multi-turn questions. They 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 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.
"""
    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 update_elo_components(max_num_files, elo_results_file):
    log_files = get_log_files(max_num_files)

    # Leaderboard
    if elo_results_file is None:  # Do live update
        battles = clean_battle_data(log_files)
        elo_results = report_elo_analysis_results(battles)

        leader_component_values[0] = make_leaderboard_md_live(elo_results)
        leader_component_values[1] = elo_results["win_fraction_heatmap"]
        leader_component_values[2] = elo_results["battle_count_heatmap"]
        leader_component_values[3] = elo_results["bootstrap_elo_rating"]
        leader_component_values[4] = elo_results["average_win_rate_bar"]

    # Basic stats
    basic_stats = report_basic_stats(log_files)
    md0 = f"Last updated: {basic_stats['last_updated_datetime']}"

    md1 = "### Action Histogram\n"
    md1 += basic_stats["action_hist_md"] + "\n"

    md2 = "### Anony. Vote Histogram\n"
    md2 += basic_stats["anony_vote_hist_md"] + "\n"

    md3 = "### Model Call Histogram\n"
    md3 += basic_stats["model_hist_md"] + "\n"

    md4 = "### Model Call (Last 24 Hours)\n"
    md4 += basic_stats["num_chats_last_24_hours"] + "\n"

    basic_component_values[0] = md0
    basic_component_values[1] = basic_stats["chat_dates_bar"]
    basic_component_values[2] = md1
    basic_component_values[3] = md2
    basic_component_values[4] = md3
    basic_component_values[5] = md4


def update_worker(max_num_files, interval, elo_results_file):
    while True:
        tic = time.time()
        update_elo_components(max_num_files, elo_results_file)
        durtaion = time.time() - tic
        print(f"update duration: {durtaion:.2f} s")
        time.sleep(max(interval - durtaion, 0))


def load_demo(url_params, request: gr.Request):
    logger.info(f"load_demo. ip: {request.client.host}. params: {url_params}")
    return basic_component_values + leader_component_values


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):
    lines = open(filename).readlines()
    heads = [v.strip() for v in lines[0].split(",")]
    rows = []
    for i in range(1, len(lines)):
        row = [v.strip() for v in lines[i].split(",")]
        for j in range(len(heads)):
            item = {}
            for h, v in zip(heads, row):
                if h == "Arena Elo rating":
                    if v != "-":
                        v = int(ast.literal_eval(v))
                    else:
                        v = np.nan
                elif h == "MMLU":
                    if v != "-":
                        v = round(ast.literal_eval(v) * 100, 1)
                    else:
                        v = np.nan
                elif h == "MT-bench (win rate %)":
                    if v != "-":
                        v = round(ast.literal_eval(v[:-1]), 1)
                    else:
                        v = np.nan
                elif h == "MT-bench (score)":
                    if v != "-":
                        v = round(ast.literal_eval(v), 2)
                    else:
                        v = np.nan
                item[h] = v
            if add_hyperlink:
                item["Model"] = model_hyperlink(item["Model"], item["Link"])
        rows.append(item)

    return rows


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(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 arena_df.index:
            idx = arena_df.index.get_loc(model_key)
            row.append(round(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]["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=["final_ranking", "rating"], ascending=[True, 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
        ranking = arena_df.iloc[i].get("final_ranking") or i+1
        row.append(ranking)
        # model display name
        row.append(model_name)
        # elo rating
        if pd.isna(arena_df.iloc[i]["rating"]):
            continue
        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}")
        # Avg. Win Rate
        row.append(f'{round(arena_df.iloc[i]["avg_win_rate_no_tie"] * 100, 1):04.1f}%')
        # num battles
        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]
        )

        #cutoff_date = model_table_df[model_table_df["key"] == model_key]["Knowledge cutoff date"].values[0]
        #if cutoff_date == "-":
        #    row.append("Unknown")
        #else:
        #    row.append(cutoff_date)

        values.append(row)
    return values

def create_leaderboard_from_results(elo_results, model_table_df, show_plot, show_language_plot=False):
    p0 = elo_results["inferred_languages_bar"]
    p1 = elo_results["win_fraction_heatmap"]
    p2 = elo_results["battle_count_heatmap"]
    p3 = elo_results["bootstrap_elo_rating"]
    p4 = elo_results["average_win_rate_bar"]
    arena_df = elo_results["leaderboard_table_df"]
    arena_table_vals = get_arena_table(arena_df, model_table_df)
    
    md = make_arena_leaderboard_md(arena_df)
    gr.Markdown(md, elem_id="leaderboard_markdown")
    gr.Dataframe(
        headers=[
            "Rank",
            "🤖 Model",
            "⭐ Arena Elo",
            "📊 95% CI",
            "🏆 Avg. Win Rate",
            "🗳️ Votes",
            "Organization",
            "License",
            #"Knowledge Cutoff",
        ],
        datatype=[
            "str",
            "markdown",
            "number",
            "str",
            "str",
            "number",
            "str",
            "str",
            #"str",
        ],
        value=arena_table_vals,
        elem_id="arena_leaderboard_dataframe",
        height=700,
        column_widths=[50, 200, 120, 100, 150, 100, 125, 125],#, 100],
        wrap=True,
    )

    gr.Markdown(
            f"""Note¹: we take the 95% confidence interval into account when determining a model's ranking.
    A model is ranked higher only if its lower bound of model score is higher than the upper bound of the other model's score. See Figure {3+int(show_language_plot)} below for visualization of the confidence intervals.  
    Note²: The Average Win Rate is calculated by assuming uniform sampling and no ties.
    """,
            elem_id="leaderboard_markdown"
        )

    if not show_plot:
        gr.Markdown(
            f""" ## Visit our [HF space]({original_leaderboard_link}) for more analysis!
            If you want to see more models, please help us [add them](https://github.com/lm-sys/FastChat/blob/main/docs/arena.md#how-to-add-a-new-model).
            """,
            elem_id="leaderboard_markdown",
        )
    else:
        gr.Markdown(
                    f"""## More Statistics for Chatbot Arena\n
        Below are figures for more statistics. The code for generating them is also included in this [notebook]({notebook_url}).
        You can find more discussions in this blog [post](https://lmsys.org/blog/2023-12-07-leaderboard/).
            """,
            elem_id="leaderboard_markdown"
        )

        show_plot_btn = gr.Button("Show plots")
        fig_id = 1
        if show_language_plot:
            gr.Markdown(
                f"#### Figure {fig_id}: Battle counts for the Top 15 Languages"
            )
            plot_0 = gr.Plot()
            fig_id += 1
        with gr.Row():
            with gr.Column():
                gr.Markdown(
                    f"#### Figure {fig_id}: Fraction of Model A Wins for All Non-tied A vs. B Battles"
                )
                plot_1 = gr.Plot()
                fig_id += 1
            with gr.Column():
                gr.Markdown(
                    f"#### Figure {fig_id}: Battle Count for Each Combination of Models (without Ties)"
                )
                plot_2 = gr.Plot()
                fig_id += 1
        with gr.Row():
            with gr.Column():
                gr.Markdown(
                    f"#### Figure {fig_id}: Confidence Intervals on Model Strength (via Bootstrapping)"
                )
                plot_3 = gr.Plot()
                fig_id += 1
            with gr.Column():
                gr.Markdown(
                    f"#### Figure {fig_id}: Average Win Rate Against All Other Models (Assuming Uniform Sampling and No Ties)"
                )
                plot_4 = gr.Plot()
                fig_id += 1

            def get_plots(*args):
                if show_language_plot:
                    return p0, p1, p2, p3, p4
                else:
                    return p1, p2, p3, p4
                
            if show_language_plot:
                show_plot_btn.click(fn=get_plots, outputs=[plot_0, plot_1, plot_2, plot_3, plot_4])
            else:
                show_plot_btn.click(fn=get_plots, outputs=[plot_1, plot_2, plot_3, plot_4])

    return p1, p2, p3, p4, plot_1, plot_2, plot_3, plot_4

def build_leaderboard_tab(elo_results_file, leaderboard_table_file, show_plot=False):
    if elo_results_file is None:  # Do live update
        default_md = "Loading ..."
        p1 = p2 = p3 = p4 = None
    else:
        with open(elo_results_file, "rb") as fin:
            elo_results = pickle.load(fin)
            #if "non-english" in elo_results:
            #    elo_results = elo_results["non-english"]

    languages = [lang for lang in elo_results if lang not in ["non-english", "full"]]
    languages = languages[::-1][:-3]
    languages_names = [lang[0].upper() + lang[1:] for lang in languages]

    default_md = make_default_md(languages_names)
    md_1 = gr.Markdown(default_md, elem_id="leaderboard_markdown")

    if leaderboard_table_file:
        data = load_leaderboard_table_csv(leaderboard_table_file)
        model_table_df = pd.DataFrame(data)

        with gr.Tabs() as tabs:
            # arena table
            with gr.Tab("Multilingual (Non-English)", id=0):
                gr.Markdown("This section includes metrics for all interactions that are not in English. See Figure 1 below for the distribution of evaluated languages.")
                p1, p2, p3, p4, plot_1, plot_2, plot_3, plot_4 = create_leaderboard_from_results(elo_results["non-english"], model_table_df, show_plot, show_language_plot=True)
            with gr.Tab("Multilingual (All langs)", id=1):
                gr.Markdown(f"This section includes metrics for all interactions, should be the same as the original [🏆LMSYS Chatbot Arena Leaderboard]({original_leaderboard_link}). See Figure 1 below for the distribution of evaluated languages.")
                create_leaderboard_from_results(elo_results['full'], model_table_df, show_plot, show_language_plot=True)
            with gr.Tab("By Language", id=2):
                with gr.Tabs() as tabs:
                    for i, lang in enumerate(languages):
                        elo_result = elo_results[lang]
                        lang = lang[0].upper() + lang[1:]
                        arena_df = elo_result['leaderboard_table_df']
                        size = round((sum(arena_df['num_battles']) // 2) / 1000)
                        with gr.Tab(lang + f" ({size}K)", id=i+3):
                            gr.Markdown(f"This section includes metrics for all interactions that are in {lang}.")
                            create_leaderboard_from_results(elo_result, model_table_df, show_plot)
            
    else:
        pass

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

    with gr.Accordion(
        "📝 Citation",
        open=True,
    ):
        citation_md = """
        ### Citation
        Please cite the following paper if you find the leaderboard or dataset helpful.
        ```
        @misc{chiang2024chatbot,
            title={Chatbot Arena: An Open Platform for Evaluating LLMs by Human Preference},
            author={Wei-Lin Chiang and Lianmin Zheng and Ying Sheng and Anastasios Nikolas Angelopoulos and Tianle Li and Dacheng Li and Hao Zhang and Banghua Zhu and Michael Jordan and Joseph E. Gonzalez and Ion Stoica},
            year={2024},
            eprint={2403.04132},
            archivePrefix={arXiv},
            primaryClass={cs.AI}
        }
        """
        gr.Markdown(citation_md, elem_id="leaderboard_markdown")
        gr.Markdown(acknowledgment_md)

    if show_plot:
        return [md_1, plot_1, plot_2, plot_3, plot_4]
    return [md_1]

block_css = """
#notice_markdown {
    font-size: 104%
}
#notice_markdown th {
    display: none;
}
#notice_markdown td {
    padding-top: 6px;
    padding-bottom: 6px;
}
#leaderboard_markdown {
    font-size: 104%
}
#leaderboard_markdown td {
    padding-top: 6px;
    padding-bottom: 6px;
}
#leaderboard_dataframe td {
    line-height: 0.1em;
}
footer {
    display:none !important
}
.sponsor-image-about img {
    margin: 0 20px;
    margin-top: 20px;
    height: 40px;
    max-height: 100%;
    width: auto;
    float: left;
}
"""

acknowledgment_md = f"""
### Acknowledgment
Thanks to LMSYS team for providing the open-source [data]({original_notebook_url}) and the original [🏆LMSYS Chatbot Arena Leaderboard]({original_leaderboard_link}).
"""

'''
def build_demo(elo_results_file, leaderboard_table_file):
    text_size = gr.themes.sizes.text_lg

    with gr.Blocks(
        title="Chatbot Arena Leaderboard",
        theme=gr.themes.Base(text_size=text_size),
        css=block_css,
    ) as demo:
        leader_components = build_leaderboard_tab(
            elo_results_file, leaderboard_table_file, show_plot=True
        )
    return demo
'''

elo_result_files = glob.glob("elo_results_*.pkl")
elo_result_files.sort(key=lambda x: int(x[12:-4]))
elo_result_file = elo_result_files[-1]

date_last_file = datetime.strptime(elo_result_file[12:-4], '%Y%m%d')

leaderboard_table_files = glob.glob("leaderboard_table_*.csv")
leaderboard_table_files.sort(key=lambda x: int(x[18:-4]))
leaderboard_table_file = leaderboard_table_files[-1]

text_size = gr.themes.sizes.text_lg

with gr.Blocks(
    title="Chatbot Arena Leaderboard",
    theme=gr.themes.Base(text_size=text_size),
    css=block_css,
) as demo:
    leader_components = build_leaderboard_tab(
        elo_result_file, leaderboard_table_file, show_plot=True
    )

if __name__ == "__main__":
    threading.Thread(target=run_pr_worker).start()
    demo.launch()