app.py

import gradio as gr
import pandas as pd
from pathlib import Path

from src.json_leaderboard import (
    create_leaderboard_df,
    get_leaderboard_stats,
    SIZE_BANDS,
    MODALITIES,
)
from src.about import (
    CITATION_BUTTON_LABEL,
    CITATION_BUTTON_TEXT,
    INTRODUCTION_TEXT,
    LLM_BENCHMARKS_TEXT,
    LINKS_AND_INFO,
    TITLE,
)
from src.display.css_html_js import custom_css


JSON_PATH = str(Path(__file__).parent / "leaderboard_data.json")
TABLE_DATATYPES = ["html", "str", "html", "str", "str", "str", "str"]


def filter_leaderboard(open_source_only, size_band, modality):
    df = create_leaderboard_df(
        JSON_PATH,
        open_source_only=open_source_only,
        size_band=size_band,
        modality=modality,
    )
    if df.empty:
        return pd.DataFrame({"Result": ["No models match the selected filters."]})
    return df


def get_stats_display():
    """Get formatted statistics for display"""
    stats = get_leaderboard_stats(JSON_PATH)

    if not stats:
        return "No statistics available"

    return f"""
    ### 📊 Leaderboard Statistics
    - **Total Models**: {stats['total_models']}
    - **Best Score**: {stats['max_acc']:.1f}
    - **Lowest Score**: {stats['min_acc']:.1f}
    """


# Create the Gradio interface
demo = gr.Blocks(css=custom_css, title="MMLongBench-Doc Leaderboard")

with demo:
    gr.HTML(TITLE)
    gr.HTML(LINKS_AND_INFO)
    gr.Markdown(INTRODUCTION_TEXT, elem_classes="markdown-text")

    with gr.Tabs(elem_classes="tab-buttons") as tabs:
        with gr.TabItem("🏅 Leaderboard", elem_id="llm-benchmark-tab-table", id=0):
            stats_display = gr.Markdown(get_stats_display())

            with gr.Row():
                open_source_filter = gr.Checkbox(
                    label="Open-source only",
                    value=False,
                )
                size_filter = gr.Dropdown(
                    choices=SIZE_BANDS,
                    value="All",
                    label="Model size (activated params)",
                )
                modality_filter = gr.Radio(
                    choices=MODALITIES,
                    value="All",
                    label="Modality",
                )

            leaderboard_table = gr.Dataframe(
                value=filter_leaderboard(False, "All", "All"),
                datatype=TABLE_DATATYPES,
                interactive=False,
                wrap=True,
            )

            for control in (open_source_filter, size_filter, modality_filter):
                control.change(
                    fn=filter_leaderboard,
                    inputs=[open_source_filter, size_filter, modality_filter],
                    outputs=leaderboard_table,
                )

    with gr.Row():
        with gr.Column():
            gr.Markdown("## 📙 Citation")
            gr.Markdown("If you use [MMLongBench-Doc](https://arxiv.org/abs/2407.01523) in your research, please cite our work:")
            citation_textbox = gr.Textbox(
                value=CITATION_BUTTON_TEXT,
                elem_id="citation-textbox",
                show_label=False,
                interactive=False,
                lines=8,
                show_copy_button=True
            )


if __name__ == "__main__":
    demo.launch()