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 gradio as gr
import numpy as np
import pandas as pd
import os
from collections import defaultdict
from matplotlib.colors import LinearSegmentedColormap

def make_default_md():
    leaderboard_md = f"""
# 🔎📚🪡📚❓ BABILong Leaderboard 🏆

[![Dataset on HF](https://huggingface.co/datasets/huggingface/badges/resolve/main/dataset-on-hf-lg.svg)](https://huggingface.co/datasets/booydar/babilong)

| [GitHub](https://github.com/booydar/babilong) | [Paper](https://arxiv.org/abs/2406.10149) | [HF Dataset](https://huggingface.co/datasets/booydar/babilong) | [HF Dataset 1k samples per task](https://huggingface.co/datasets/RMT-team/babilong-1k-samples) |
"""
    return leaderboard_md

def make_arena_leaderboard_md(total_models):
    leaderboard_md = f"""Total #models: **{total_models}**. Last updated: July 29, 2024."""
    return leaderboard_md

def make_model_desc_md(f_len):
    desc_md = make_arena_leaderboard_md(f_len)
    models = next(os.walk('info'))[2]
    for model in models:
        model_name = model.split('.md')[0]
        with open(os.path.join('info', model), 'r') as f:
            description = f.read()
        desc_md += f"\n\n### {model_name}\n{description}"
    return desc_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_model(folders, tab_name, msg_lengths):
    results = defaultdict(list)

    class NA():
        def __repr__(self) -> str:
            return '-'
        def __float__(self):
            return 0.0
       
    mean_score = []

    for i, folder in enumerate(folders):
        model_name = folder.split('/')[-1]
        if 'fine-tune' in model_name:
            model_name += ' 🛠️'
        if 'rag' in model_name.lower() or 'retrieve' in model_name.lower():
            model_name += ' 🔎'
        results['Model'].append(model_name)
        for task in msg_lengths:
            if not os.path.isfile(f'{folder}/{tab_name}/{task}.csv'):
                results[msg_lengths[task]].append(NA())
            else:
                df = pd.read_csv(f'{folder}/{tab_name}/{task}.csv')
                results[msg_lengths[task]].append(int(df['result'].sum() / len(df) * 100))

        mean_score.append(-np.mean([float(results[msg_lengths[task]][i]) for task in list(msg_lengths.keys())[:10]]))

    res_df = pd.DataFrame(results)
    lengths = list(msg_lengths.values())
    res_df['mean_score'] = mean_score
    res_df['num_lengths'] = -(res_df[lengths].astype(float) > 0).sum(axis=1)
    res_df = res_df[res_df.num_lengths != 0]
    res_df.sort_values(['num_lengths', 'mean_score'], inplace=True)
    res_df['Rank'] = range(1, res_df.shape[0] + 1)

    res_df['Avg ≤32k'] = res_df[lengths[:5]].astype(float).fillna(0).mean(axis=1).astype(int)
    res_df['Avg ≤128k'] = res_df[lengths[:7]].astype(float).fillna(0).mean(axis=1).astype(int)
    ordered_columns = ['Rank', 'Model', 'Avg ≤32k', 'Avg ≤128k'] + lengths
    res_df = res_df[ordered_columns]
    return res_df
    
def build_leaderboard_tab(folders):
    default_md = make_default_md()
    md_1 = gr.Markdown(default_md, elem_id="leaderboard_markdown")
    msg_lengths = {
        '0': '0k', 
        '4000': '4k', 
        '8000': '8k', 
        '16000': '16k', 
        '32000': '32k', 
        '64000': '64k', 
        '128000': '128k', 
        '500000': '500k', 
        '1000000': '1M', 
        '10000000': '10M'
    }

    with gr.Tabs() as tabs:
        for tab_id, tab_name in enumerate(['avg', 'qa1','qa2', 'qa3', 'qa4', 'qa5']):
                df = load_model(folders, tab_name, msg_lengths)
                cmap = LinearSegmentedColormap.from_list('ryg', ["red", "yellow", "green"], N=256)

                # df = df.style.background_gradient(cmap=cmap, vmin=0, vmax=100, subset=list(msg_lengths.values()))
                df = df.style.background_gradient(cmap=cmap, vmin=0, vmax=100, subset=df.columns[2:])
                # arena table
                with gr.Tab(tab_name, id=tab_id):
                    md = make_arena_leaderboard_md(len(folders))
                    gr.Markdown(md, elem_id="leaderboard_markdown")
                    gr.Dataframe(
                        headers=[
                            "Rank",
                            "Model",
                        ] + list(msg_lengths.values()) + ['Avg ≤32k', 'Avg ≤128k'],
                        datatype=[
                            "str",
                            "markdown",
                            "str",
                            "str",
                            "str",
                            "str",
                            "str",
                            "str",
                            "str",
                            "str",
                            "str",
                        ],
                        value=df,
                        elem_id="arena_leaderboard_dataframe",
                        height=700,
                        column_widths=[20, 150] + [30] * 2 + [20] * len(msg_lengths),
                        wrap=True,
                    )

        with gr.Tab("Description", id=tab_id + 1):
            desc_md = make_model_desc_md(len(folders))
            gr.Markdown(desc_md, elem_id="leaderboard_markdown")

    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
}
.image-container {
    display: flex;
    align-items: center;
    padding: 1px;
}
.image-container img {
    margin: 0 30px;
    height: 20px;
    max-height: 100%;
    width: auto;
    max-width: 20%;
}
"""

def build_demo(folders):
    text_size = gr.themes.sizes.text_lg

    with gr.Blocks(
        title="Babilong leaderboard",
        theme=gr.themes.Base(text_size=text_size),
        css=block_css,
    ) as demo:
        leader_components = build_leaderboard_tab(folders)
    return demo


if __name__ == "__main__":
    folders = [f'results/{folders}' for folders in os.listdir('results')]
    demo = build_demo(folders)
    demo.launch(share=False)