app.py

import gradio as gr
import pandas as pd

from dataset import get_dataframe
from markdown import GUIDELINES, PANEL_MARKDOWN


df = get_dataframe()


def filter_dataframe(dataframe, eval_dataset, cont_source, checkboxes):
    """
    Filter the dataframe based on the provided evaluation dataset, contaminated source, and checkboxes.

    Args:
        dataframe (pandas.DataFrame): The input dataframe to filter.
        eval_dataset (str): The evaluation dataset to filter by.
        cont_source (str): The contaminated source to filter by.
        checkboxes (list): The checkboxes to filter by.

    Returns:
        pandas.DataFrame: The filtered dataframe.
    """
    if isinstance(eval_dataset, str):
        dataframe = dataframe[
            dataframe["Evaluation Dataset"].str.contains(eval_dataset)
        ]
    if isinstance(cont_source, str):
        dataframe = dataframe[
            dataframe["Contaminated Source"].str.contains(cont_source)
        ]
    if isinstance(checkboxes, list) and "Exclude model-based evidences" in checkboxes:
        dataframe = dataframe[dataframe["Approach"] != "model-based"]
    if isinstance(checkboxes, list) and "Show only contaminated" in checkboxes:
        dataframe = dataframe[
            (dataframe["Train Split"] > 0.0)
            | (dataframe["Development Split"] > 0.0)
            | (dataframe["Test Split"] > 0.0)
        ]

    return dataframe.style.format(
        {"Train Split": "{:.1%}", "Development Split": "{:.1%}", "Test Split": "{:.1%}"}
    )


def filter_dataframe_corpus(*args, **kwargs) -> pd.DataFrame:
    """
    Filter the dataframe for corpus contamination.

    Returns:
        pandas.DataFrame: The filtered dataframe for corpus contamination.
    """
    # Get rows in which the column Model or corpus is equal to dataset
    filtered_df = df[df["Model or corpus"] == "corpus"]
    filtered_df = filtered_df.drop(columns=["Model or corpus"])
    return filter_dataframe(filtered_df, *args, **kwargs)


def filter_dataframe_model(*args, **kwargs) -> pd.DataFrame:
    """
    Filter the dataframe for model contamination.

    Returns:
        pandas.DataFrame: The filtered dataframe for model contamination.
    """
    # Get rows in which the column Model or corpus is equal to dataset
    filtered_df = df[df["Model or corpus"] == "model"]
    filtered_df = filtered_df.drop(columns=["Model or corpus"])
    return filter_dataframe(filtered_df, *args, **kwargs)


theme = gr.themes.Soft(
    primary_hue="emerald",
    secondary_hue="red",
    text_size="sm",
    spacing_size="sm",
    font=[
        gr.themes.GoogleFont("Poppins"),
        gr.themes.GoogleFont("Poppins"),
        gr.themes.GoogleFont("Poppins"),
        gr.themes.GoogleFont("Poppins"),
    ],
).set(block_background_fill="*neutral_50", block_background_fill_dark="*neutral_950")


with gr.Blocks(
    theme=theme,
    title="💨 Data Contamination Report",
    analytics_enabled=False,
    fill_height=True,
) as demo:
    gr.Markdown(PANEL_MARKDOWN)
    with gr.Tab("Corpus contamination") as tab_corpus:
        with gr.Row(variant="compact"):
            with gr.Column():
                eval_dataset_corpus = gr.Textbox(
                    placeholder="Evaluation dataset",
                    label="Evaluation dataset",
                    value="",
                )
                cont_corpora = gr.Textbox(
                    placeholder="Pre-training corpora",
                    label="Pre-training corpora",
                    value="",
                )
            with gr.Column():
                checkboxes_corpus = gr.CheckboxGroup(
                    ["Exclude model-based evidences", "Show only contaminated"],
                    label="Search options",
                    value=[],
                )

        filter_corpus_btn = gr.Button("Filter")

        corpus_dataframe = gr.DataFrame(
            value=filter_dataframe_corpus(
                eval_dataset_corpus, cont_corpora, checkboxes_corpus
            ),
            headers=df.columns.to_list(),
            datatype=[
                "markdown",
                "markdown",
                "number",
                "number",
                "number",
                "str",
                "markdown",
                "markdown",
            ],
        )

    with gr.Tab("Model contamination") as tab_model:
        with gr.Row(variant="compact"):
            with gr.Column():
                eval_dataset_model = gr.Textbox(
                    placeholder="Evaluation dataset",
                    label="Evaluation dataset",
                    value="",
                )
                cont_model = gr.Textbox(
                    placeholder="Model", label="Pre-training corpora", value=""
                )
            with gr.Column():
                checkboxes_model = gr.CheckboxGroup(
                    ["Exclude model-based evidences", "Show only contaminated"],
                    label="Search options",
                    value=[],
                )

        filter_model_btn = gr.Button("Filter")

        model_dataframe = gr.DataFrame(
            value=filter_dataframe_model(
                eval_dataset_model, cont_model, checkboxes_model
            ),
            headers=df.columns.to_list(),
            datatype=[
                "markdown",
                "markdown",
                "number",
                "number",
                "number",
                "str",
                "markdown",
                "markdown",
            ],
        )

    filter_corpus_btn.click(
        filter_dataframe_corpus,
        inputs=[eval_dataset_corpus, cont_corpora, checkboxes_corpus],
        outputs=corpus_dataframe,
    )
    filter_model_btn.click(
        filter_dataframe_model,
        inputs=[eval_dataset_model, cont_model, checkboxes_model],
        outputs=model_dataframe,
    )

    with gr.Tab("Contribution Guidelines") as tab_guidelines:
        gr.Markdown(GUIDELINES)


demo.launch()