app.py

import os
import traceback

import gradio as gr
import pandas as pd
import plotly.express as px
import plotly.graph_objects as go
from apscheduler.schedulers.background import BackgroundScheduler
from huggingface_hub import snapshot_download

from src.about import (
    BOTTOM_LOGO,
    CITATION_BUTTON_LABEL,
    CITATION_BUTTON_LABEL_JA,
    CITATION_BUTTON_TEXT,
    EVALUATION_QUEUE_TEXT,
    EVALUATION_QUEUE_TEXT_JA,
    INTRODUCTION_TEXT,
    INTRODUCTION_TEXT_JA,
    LLM_BENCHMARKS_TEXT,
    LLM_BENCHMARKS_TEXT_JA,
    TITLE,
    TaskType,
)
from src.display.utils import (
    BENCHMARK_COLS,
    COLS,
    EVAL_COLS,
    EVAL_TYPES,
    NUMERIC_INTERVALS,
    TYPES,
    AddSpecialTokens,
    ApplyChatTemplate,
    AutoEvalColumn,
    EnableThinking,
    ModelType,
    Precision,
    fields,
)
from src.envs import API, CONTENTS_REPO, EVAL_REQUESTS_PATH, QUEUE_REPO, REPO_ID
from src.i18n import (
    CITATION_ACCORDION_LABEL,
    CITATION_ACCORDION_LABEL_JA,
    SELECT_ALL_BUTTON_LABEL,
    SELECT_ALL_BUTTON_LABEL_JA,
    SELECT_AVG_ONLY_BUTTON_LABEL,
    SELECT_AVG_ONLY_BUTTON_LABEL_JA,
    SELECT_NONE_BUTTON_LABEL,
    SELECT_NONE_BUTTON_LABEL_JA,
)
from src.populate import get_evaluation_queue_df, get_leaderboard_df
from src.submission.submit import cancel_eval, confirm_eval, preview_eval


def restart_space() -> None:
    API.restart_space(repo_id=REPO_ID)


# Queue sync is best-effort — the leaderboard can still serve without it.
try:
    snapshot_download(
        repo_id=QUEUE_REPO,
        local_dir=EVAL_REQUESTS_PATH,
        repo_type="dataset",
        tqdm_class=None,
        etag_timeout=30,
        token=API.token,
    )
except Exception as e:
    print(
        f"[startup] Failed to sync eval queue from {QUEUE_REPO}: "
        f"{type(e).__name__}: {e}",
        flush=True,
    )
    traceback.print_exc()
    EVAL_REQUESTS_PATH.mkdir(parents=True, exist_ok=True)


# Get dataframes

try:
    (
        FINISHED_EVAL_QUEUE_DF,
        RUNNING_EVAL_QUEUE_DF,
        PENDING_EVAL_QUEUE_DF,
        FAILED_EVAL_QUEUE_DF,
    ) = get_evaluation_queue_df(EVAL_REQUESTS_PATH, EVAL_COLS)
except Exception as e:
    print(
        f"[startup] Failed to build eval queue df from {EVAL_REQUESTS_PATH}: "
        f"{type(e).__name__}: {e}",
        flush=True,
    )
    traceback.print_exc()
    _empty_queue_df = pd.DataFrame(columns=EVAL_COLS)
    FINISHED_EVAL_QUEUE_DF = _empty_queue_df
    RUNNING_EVAL_QUEUE_DF = _empty_queue_df
    PENDING_EVAL_QUEUE_DF = _empty_queue_df
    FAILED_EVAL_QUEUE_DF = _empty_queue_df

try:
    ORIGINAL_DF = get_leaderboard_df(CONTENTS_REPO, COLS, BENCHMARK_COLS)
except Exception as e:
    print(
        f"[startup] Failed to load leaderboard from {CONTENTS_REPO}: "
        f"{type(e).__name__}: {e}",
        flush=True,
    )
    traceback.print_exc()
    ORIGINAL_DF = pd.DataFrame()


# Searching and filtering


def filter_models(
    df: pd.DataFrame,
    type_query: list[str],
    size_query: list[str],
    precision_query: list[str],
    add_special_tokens_query: list[str],
    enable_thinking_query: list[str],
    apply_chat_template_query: list[str],
) -> pd.DataFrame:
    # Filter by model type
    type_emoji = [t.split()[0] for t in type_query]
    df = df[df["T"].isin(type_emoji)]

    # Filter by precision
    df = df[df["Precision"].isin(precision_query)]

    # Filter by model size
    # Note: When `df` is empty, `size_mask` is empty, and the shape of `df[size_mask]` becomes (0, 0),
    # so we need to check the length of `df` before applying the filter.
    if len(df) > 0:
        size_mask = df["#Params (B)"].apply(
            lambda x: any(x in NUMERIC_INTERVALS[s] for s in size_query if s != "Unknown")
        )
        if "Unknown" in size_query:
            size_mask |= df["#Params (B)"].isna() | (df["#Params (B)"] == 0)
        df = df[size_mask]

    # Filter by special tokens setting
    df = df[df["Add Special Tokens"].isin(add_special_tokens_query)]

    # Filter by enable_thinking
    df = df[df["Enable Thinking"].isin(enable_thinking_query)]

    # Filter by apply_chat_template
    df = df[df["Apply Chat Template"].isin(apply_chat_template_query)]

    return df


def search_model_by_name(df: pd.DataFrame, model_name: str) -> pd.DataFrame:
    return df[df[AutoEvalColumn.dummy.name].str.contains(model_name, case=False)]


def search_models_by_multiple_names(df: pd.DataFrame, search_text: str) -> pd.DataFrame:
    if not search_text:
        return df
    model_names = [name.strip() for name in search_text.split(";")]
    dfs = [search_model_by_name(df, name) for name in model_names if name]
    return pd.concat(dfs).drop_duplicates(subset=AutoEvalColumn.row_id.name)


def select_columns(df: pd.DataFrame, columns: list[str]) -> pd.DataFrame:
    always_here_cols = [
        AutoEvalColumn.model_type_symbol.name,  # 'T'
        AutoEvalColumn.model.name,  # 'Model'
    ]

    # Remove 'always_here_cols' from 'columns' to avoid duplicates
    columns = [c for c in columns if c not in always_here_cols]
    new_columns = (
        always_here_cols + [c for c in COLS if c in df.columns and c in columns] + [AutoEvalColumn.row_id.name]
    )

    # Maintain order while removing duplicates
    seen = set()
    unique_columns = []
    for c in new_columns:
        if c not in seen:
            unique_columns.append(c)
            seen.add(c)

    # Create DataFrame with filtered columns
    filtered_df = df[unique_columns]
    return filtered_df


def update_table(
    type_query: list[str],
    precision_query: list[str],
    size_query: list[str],
    add_special_tokens_query: list[str],
    enable_thinking_query: list[str],
    apply_chat_template_query: list[str],
    query: str,
    *columns,
) -> pd.DataFrame:
    columns = [item for column in columns for item in column]
    df = filter_models(
        ORIGINAL_DF,
        type_query,
        size_query,
        precision_query,
        add_special_tokens_query,
        enable_thinking_query,
        apply_chat_template_query,
    )
    df = search_models_by_multiple_names(df, query)
    df = select_columns(df, columns)
    return df


# Prepare the dataframes


INITIAL_COLUMNS = ["T"] + [
    c.name for c in fields(AutoEvalColumn) if (c.never_hidden or c.displayed_by_default) and c.name != "T"
]
leaderboard_df = ORIGINAL_DF.copy()
if len(leaderboard_df) > 0:
    leaderboard_df = filter_models(
        leaderboard_df,
        [t.to_str(" : ") for t in ModelType],
        list(NUMERIC_INTERVALS.keys()),
        [i.value.name for i in Precision],
        [i.value.name for i in AddSpecialTokens],
        [i.value.name for i in EnableThinking],
        [i.value.name for i in ApplyChatTemplate],
    )
    leaderboard_df = select_columns(leaderboard_df, INITIAL_COLUMNS)
else:
    leaderboard_df = pd.DataFrame(columns=INITIAL_COLUMNS)

# Leaderboard demo


def toggle_all_categories(action: str) -> list[gr.CheckboxGroup]:
    """Function to control all category checkboxes at once"""
    results = []
    for task_type in TaskType:
        if task_type == TaskType.NotTask:
            # Maintain existing selection for Model details
            results.append(gr.CheckboxGroup())
        elif action == "all":
            # Select all
            results.append(
                gr.CheckboxGroup(
                    value=[
                        c.name
                        for c in fields(AutoEvalColumn)
                        if not c.hidden and not c.never_hidden and not c.dummy and c.task_type == task_type
                    ]
                )
            )
        elif action == "none":
            # Deselect all
            results.append(gr.CheckboxGroup(value=[]))
        elif action == "avg_only":
            # Select only AVG metrics
            results.append(
                gr.CheckboxGroup(
                    value=[
                        c.name
                        for c in fields(AutoEvalColumn)
                        if not c.hidden
                        and not c.never_hidden
                        and c.task_type == task_type
                        and ((task_type == TaskType.AVG) or (task_type != TaskType.AVG and c.average))
                    ]
                )
            )
    return results


TASK_AVG_NAME_MAP = {
    c.name: c.task_type.name for c in fields(AutoEvalColumn) if c.average and c.task_type != TaskType.AVG
}
AVG_COLUMNS = ["AVG"] + list(TASK_AVG_NAME_MAP.keys())


def plot_size_vs_score(df_filtered: pd.DataFrame) -> go.Figure:
    if len(ORIGINAL_DF) == 0 or AutoEvalColumn.row_id.name not in ORIGINAL_DF.columns:
        return go.Figure()
    df = ORIGINAL_DF[ORIGINAL_DF[AutoEvalColumn.row_id.name].isin(df_filtered[AutoEvalColumn.row_id.name])]
    df = df[df["#Params (B)"] > 0]
    available_avg = [c for c in AVG_COLUMNS if c in df.columns]
    df = df[["model_name_for_query", "#Params (B)"] + available_avg]
    df = df.rename(columns={"model_name_for_query": "Model"})
    df["model_name_without_org_name"] = df["Model"].str.split("/").str[-1]
    df = pd.melt(
        df,
        id_vars=["Model", "model_name_without_org_name", "#Params (B)"],
        value_vars=available_avg,
        var_name="Category",
        value_name="Score",
    )
    max_model_size = df["#Params (B)"].max()
    fig = px.scatter(
        df,
        x="#Params (B)",
        y="Score",
        text="model_name_without_org_name",
        color="Category",
        hover_data=["Model", "Category"],
    )
    fig.update_traces(
        hovertemplate="<b>%{customdata[0]}</b><br>#Params: %{x:.2f}B<br>%{customdata[1]}: %{y:.4f}<extra></extra>",
        textposition="top right",
        mode="markers+text",
    )
    for trace in fig.data:
        if trace.name != "AVG":
            trace.visible = "legendonly"
    fig.update_layout(xaxis_range=[0, max_model_size * 1.2], yaxis_range=[0, 1])
    fig.update_layout(
        updatemenus=[
            {
                "type": "buttons",
                "direction": "left",
                "showactive": True,
                "buttons": [
                    {"label": "Hide Labels", "method": "restyle", "args": ["mode", "markers"]},
                    {"label": "Show Labels", "method": "restyle", "args": ["mode", "markers+text"]},
                ],
                "x": 0.5,
                "y": -0.2,
                "xanchor": "center",
                "yanchor": "top",
            }
        ]
    )
    return fig


def plot_average_scores(df_filtered: pd.DataFrame) -> go.Figure:
    if len(ORIGINAL_DF) == 0 or AutoEvalColumn.row_id.name not in ORIGINAL_DF.columns:
        return go.Figure()
    df = ORIGINAL_DF[ORIGINAL_DF[AutoEvalColumn.row_id.name].isin(df_filtered[AutoEvalColumn.row_id.name])]
    available_avg_keys = [k for k in TASK_AVG_NAME_MAP if k in df.columns]
    df = df[["model_name_for_query"] + available_avg_keys]
    df = df.rename(columns={"model_name_for_query": "Model"})
    df = df.rename(columns={k: TASK_AVG_NAME_MAP[k] for k in available_avg_keys})
    df = df.set_index("Model")

    fig = go.Figure()
    for i, (name, row) in enumerate(df.iterrows()):
        visible = True if i < 2 else "legendonly"  # Display only the first 2 models
        fig.add_trace(
            go.Scatterpolar(
                r=row.values,
                theta=row.index,
                fill="toself",
                name=name,
                hovertemplate="%{theta}: %{r}",
                visible=visible,
            )
        )
    fig.update_layout(
        polar={
            "radialaxis": {"range": [0, 1]},
        },
        showlegend=True,
    )
    return fig


shown_columns_dict: dict[str, gr.CheckboxGroup] = {}
checkboxes: list[gr.CheckboxGroup] = []

with gr.Blocks() as demo_leaderboard:
    with gr.Row():
        search_bar = gr.Textbox(
            placeholder=" 🔍 Search for your model (separate multiple queries with `;`) and press ENTER...",
            show_label=False,
            elem_id="search-bar",
        )
    with gr.Accordion("Column Filter", open=True):
        with gr.Row():
            with gr.Row():
                select_all_button = gr.Button(SELECT_ALL_BUTTON_LABEL_JA, size="sm")
                select_none_button = gr.Button(SELECT_NONE_BUTTON_LABEL_JA, size="sm")
                select_avg_only_button = gr.Button(SELECT_AVG_ONLY_BUTTON_LABEL_JA, size="sm")

            for task_type in TaskType:
                label = "Model details" if task_type == TaskType.NotTask else task_type.value
                with gr.Accordion(label, open=True, elem_classes="accordion"):
                    with gr.Row(height=110):
                        shown_column = gr.CheckboxGroup(
                            show_label=False,
                            choices=[
                                c.name
                                for c in fields(AutoEvalColumn)
                                if not c.hidden and not c.never_hidden and not c.dummy and c.task_type == task_type
                            ],
                            value=[
                                c.name
                                for c in fields(AutoEvalColumn)
                                if c.displayed_by_default
                                and not c.hidden
                                and not c.never_hidden
                                and c.task_type == task_type
                            ],
                            elem_id="column-select",
                            container=False,
                        )
                        shown_columns_dict[task_type.name] = shown_column
                        checkboxes.append(shown_column)

    with gr.Accordion("Model Filter", open=True):
        with gr.Row():
            filter_columns_type = gr.CheckboxGroup(
                label="Model types",
                choices=[t.to_str() for t in ModelType],
                value=[t.to_str() for t in ModelType],
                elem_id="filter-columns-type",
            )
            filter_columns_precision = gr.CheckboxGroup(
                label="Precision",
                choices=[i.value.name for i in Precision],
                value=[i.value.name for i in Precision],
                elem_id="filter-columns-precision",
            )
            filter_columns_size = gr.CheckboxGroup(
                label="Model sizes (in billions of parameters)",
                choices=list(NUMERIC_INTERVALS.keys()),
                value=list(NUMERIC_INTERVALS.keys()),
                elem_id="filter-columns-size",
            )
            filter_columns_add_special_tokens = gr.CheckboxGroup(
                label="Add Special Tokens",
                choices=[i.value.name for i in AddSpecialTokens],
                value=[i.value.name for i in AddSpecialTokens],
                elem_id="filter-columns-add-special-tokens",
            )
            filter_columns_enable_thinking = gr.CheckboxGroup(
                label="Enable Thinking",
                choices=[i.value.name for i in EnableThinking],
                value=[i.value.name for i in EnableThinking],
                elem_id="filter-columns-enable-thinking",
            )
            filter_columns_apply_chat_template = gr.CheckboxGroup(
                label="Apply Chat Template",
                choices=[i.value.name for i in ApplyChatTemplate],
                value=[i.value.name for i in ApplyChatTemplate],
                elem_id="filter-columns-apply-chat-template",
            )

    leaderboard_table = gr.Dataframe(
        value=leaderboard_df,
        headers=INITIAL_COLUMNS,
        datatype=TYPES,
        elem_id="leaderboard-table",
        interactive=False,
        visible=True,
    )

    graph_size_vs_score = gr.Plot(label="Size vs. Score", value=plot_size_vs_score(leaderboard_df))
    graph_average_scores = gr.Plot(
        label="Performance across Task Categories", value=plot_average_scores(leaderboard_df)
    )

    select_all_button.click(
        fn=lambda: toggle_all_categories("all"),
        outputs=checkboxes,
        api_name=False,
        queue=False,
    )
    select_none_button.click(
        fn=lambda: toggle_all_categories("none"),
        outputs=checkboxes,
        api_name=False,
        queue=False,
    )
    select_avg_only_button.click(
        fn=lambda: toggle_all_categories("avg_only"),
        outputs=checkboxes,
        api_name=False,
        queue=False,
    )

    gr.on(
        triggers=[
            filter_columns_type.change,
            filter_columns_precision.change,
            filter_columns_size.change,
            filter_columns_add_special_tokens.change,
            filter_columns_enable_thinking.change,
            filter_columns_apply_chat_template.change,
            search_bar.submit,
        ]
        + [shown_columns.change for shown_columns in shown_columns_dict.values()],
        fn=update_table,
        inputs=[
            filter_columns_type,
            filter_columns_precision,
            filter_columns_size,
            filter_columns_add_special_tokens,
            filter_columns_enable_thinking,
            filter_columns_apply_chat_template,
            search_bar,
        ]
        + list(shown_columns_dict.values()),
        outputs=leaderboard_table,
    )

    leaderboard_table.change(
        fn=plot_size_vs_score,
        inputs=leaderboard_table,
        outputs=graph_size_vs_score,
        api_name=False,
        queue=False,
    )

    leaderboard_table.change(
        fn=plot_average_scores,
        inputs=leaderboard_table,
        outputs=graph_average_scores,
        api_name=False,
        queue=False,
    )


# Submission demo


def display_user_info(profile: gr.OAuthProfile | None) -> str:
    """Display user information if logged in"""
    if profile is None:
        return "Please log in to submit a model"
    return f"Logged in as: **{profile.name}** (@{profile.username})"


with gr.Blocks() as demo_submission:
    with gr.Column():
        with gr.Row():
            evaluation_queue_text = gr.Markdown(EVALUATION_QUEUE_TEXT_JA, elem_classes="markdown-text")

        with gr.Column():
            with gr.Accordion(
                f"✅ Finished Evaluations ({len(FINISHED_EVAL_QUEUE_DF)})",
                open=False,
            ):
                with gr.Row():
                    finished_eval_table = gr.Dataframe(
                        value=FINISHED_EVAL_QUEUE_DF,
                        headers=EVAL_COLS,
                        datatype=EVAL_TYPES,
                        row_count=5,
                    )
            with gr.Accordion(
                f"🔄 Running Evaluation Queue ({len(RUNNING_EVAL_QUEUE_DF)})",
                open=True,
            ):
                with gr.Row():
                    running_eval_table = gr.Dataframe(
                        value=RUNNING_EVAL_QUEUE_DF,
                        headers=EVAL_COLS,
                        datatype=EVAL_TYPES,
                        row_count=5,
                    )

            with gr.Accordion(
                f"⏳ Pending Evaluation Queue ({len(PENDING_EVAL_QUEUE_DF)})",
                open=False,
            ):
                with gr.Row():
                    pending_eval_table = gr.Dataframe(
                        value=PENDING_EVAL_QUEUE_DF,
                        headers=EVAL_COLS,
                        datatype=EVAL_TYPES,
                        row_count=5,
                    )
            with gr.Accordion(
                f"❎ Failed Evaluation Queue ({len(FAILED_EVAL_QUEUE_DF)})",
                open=False,
            ):
                with gr.Row():
                    failed_eval_table = gr.Dataframe(
                        value=FAILED_EVAL_QUEUE_DF,
                        headers=EVAL_COLS,
                        datatype=EVAL_TYPES,
                        row_count=5,
                    )
    with gr.Row():
        with gr.Column(scale=3):
            gr.Markdown("# ✉️✨ Submit your model here!", elem_classes="markdown-text")

    with gr.Row():
        user_info_markdown = gr.Markdown("Please log in to submit a model", elem_classes="markdown-text")

    with gr.Row():
        gr.LoginButton(size="lg")

    with gr.Row():
        with gr.Column():
            model_name_textbox = gr.Textbox(label="Model name")
            revision_name_textbox = gr.Textbox(label="Revision commit", placeholder="main")
            model_type = gr.Dropdown(
                label="Model type",
                choices=[t.to_str(" : ") for t in ModelType],
                multiselect=False,
                value=None,
            )

        with gr.Column():
            precision = gr.Dropdown(
                label="Precision",
                choices=[i.value.name for i in Precision] + ["auto"],
                multiselect=False,
                value="auto",
            )
            add_special_tokens = gr.Dropdown(
                label="AddSpecialTokens",
                choices=[i.value.name for i in AddSpecialTokens],
                multiselect=False,
                value="False",
            )

    with gr.Row():
        with gr.Column():
            apply_chat_template = gr.Dropdown(
                label="Apply Chat Template",
                choices=["False", "True"],
                multiselect=False,
                value="True",
                info="Whether to apply chat template to the model",
            )
            enable_thinking = gr.Dropdown(
                label="Enable Thinking (Reasoning mode)",
                choices=["False", "True"],
                multiselect=False,
                value="False",
                info="Enable thinking mode (only for compatible models like Qwen3). Max reasoning length: 1024 tokens",
            )

        with gr.Column():
            reasoning_parser = gr.Dropdown(
                label="Reasoning Parser",
                choices=[
                    "",
                    "deepseek_r1",
                    "deepseek_v3",
                    "ernie45",
                    "gemma4",
                    "glm45",
                    "openai_gptoss",
                    "granite",
                    "hunyuan_a13b",
                    "kimi_k2",
                    "llmjp4",
                    "minimax_m2",
                    "minimax_m2_append_think",
                    "mistral",
                    "olmo3",
                    "qwen3",
                    "seed_oss",
                    "step3",
                ],
                multiselect=False,
                value="",
                info="Reasoning parser type (only effective when Enable Thinking=True)",
            )

    submit_button = gr.Button("Submit Eval")
    submission_result = gr.Markdown()
    eval_entry_state = gr.State(value=None)

    with gr.Column(visible=False, elem_classes="confirm-modal-overlay") as modal_overlay:
        with gr.Column(elem_classes="confirm-modal-box"):
            modal_content = gr.Markdown()
            with gr.Row():
                confirm_button = gr.Button("Confirm", variant="primary", scale=1)
                cancel_button = gr.Button("Cancel", variant="secondary", scale=1)

    submit_inputs = [
        model_name_textbox,
        revision_name_textbox,
        precision,
        model_type,
        add_special_tokens,
        apply_chat_template,
        enable_thinking,
        reasoning_parser,
    ]

    submit_button.click(
        fn=preview_eval,
        inputs=submit_inputs,
        outputs=[submission_result, modal_content, modal_overlay, eval_entry_state],
        api_name="preview_eval",
    )
    confirm_button.click(
        fn=confirm_eval,
        inputs=[eval_entry_state],
        outputs=[submission_result, modal_overlay, eval_entry_state],
        api_name="confirm_eval",
    )
    cancel_button.click(
        fn=cancel_eval,
        inputs=[],
        outputs=[submission_result, modal_overlay, eval_entry_state],
    )

    # Load user info when the page loads
    demo_submission.load(fn=display_user_info, outputs=user_info_markdown)


# Main demo


def set_default_language(request: gr.Request) -> gr.Radio:
    if request.headers["Accept-Language"].split(",")[0].lower().startswith("ja"):
        return gr.Radio(value="🇯🇵 JA")
    else:
        return gr.Radio(value="🇺🇸 EN")


def update_language(
    language: str,
) -> tuple[
    gr.Markdown,  # introduction_text
    gr.Markdown,  # llm_benchmarks_text
    gr.Markdown,  # evaluation_queue_text
    gr.Textbox,  # citation_button
    gr.Button,  # select_all_button
    gr.Button,  # select_none_button
    gr.Button,  # select_avg_only_button
    gr.Accordion,  # citation_accordion
]:
    if language == "🇯🇵 JA":
        return (
            gr.Markdown(value=INTRODUCTION_TEXT_JA),
            gr.Markdown(value=LLM_BENCHMARKS_TEXT_JA),
            gr.Markdown(value=EVALUATION_QUEUE_TEXT_JA),
            gr.Textbox(label=CITATION_BUTTON_LABEL_JA),
            gr.Button(value=SELECT_ALL_BUTTON_LABEL_JA),
            gr.Button(value=SELECT_NONE_BUTTON_LABEL_JA),
            gr.Button(value=SELECT_AVG_ONLY_BUTTON_LABEL_JA),
            gr.Accordion(label=CITATION_ACCORDION_LABEL_JA),
        )
    else:
        return (
            gr.Markdown(value=INTRODUCTION_TEXT),
            gr.Markdown(value=LLM_BENCHMARKS_TEXT),
            gr.Markdown(value=EVALUATION_QUEUE_TEXT),
            gr.Textbox(label=CITATION_BUTTON_LABEL),
            gr.Button(value=SELECT_ALL_BUTTON_LABEL),
            gr.Button(value=SELECT_NONE_BUTTON_LABEL),
            gr.Button(value=SELECT_AVG_ONLY_BUTTON_LABEL),
            gr.Accordion(label=CITATION_ACCORDION_LABEL),
        )


with gr.Blocks(css_paths="style.css", theme=gr.themes.Glass()) as demo:
    gr.HTML(TITLE)
    introduction_text = gr.Markdown(INTRODUCTION_TEXT_JA, elem_classes="markdown-text")

    with gr.Tabs() as tabs:
        with gr.Tab("🏅 LLM Benchmark", elem_id="llm-benchmark-tab-table"):
            demo_leaderboard.render()

        with gr.Tab("📝 About", elem_id="llm-benchmark-tab-about"):
            llm_benchmarks_text = gr.Markdown(LLM_BENCHMARKS_TEXT_JA, elem_classes="markdown-text")

        with gr.Tab("🚀 Submit here! ", elem_id="llm-benchmark-tab-submit"):
            demo_submission.render()

    with gr.Row():
        with gr.Accordion(CITATION_ACCORDION_LABEL_JA, open=False) as citation_accordion:
            citation_button = gr.Textbox(
                label=CITATION_BUTTON_LABEL_JA,
                value=CITATION_BUTTON_TEXT,
                lines=20,
                elem_id="citation-button",
                show_copy_button=True,
            )
    gr.HTML(BOTTOM_LOGO)

    language = gr.Radio(
        choices=["🇯🇵 JA", "🇺🇸 EN"],
        value="🇯🇵 JA",
        elem_classes="language-selector",
        show_label=False,
        container=False,
    )

    demo.load(fn=set_default_language, outputs=language)
    language.change(
        fn=update_language,
        inputs=language,
        outputs=[
            introduction_text,
            llm_benchmarks_text,
            evaluation_queue_text,
            citation_button,
            select_all_button,
            select_none_button,
            select_avg_only_button,
            citation_accordion,
        ],
        api_name=False,
    )

if __name__ == "__main__":
    if os.getenv("SPACE_ID"):
        scheduler = BackgroundScheduler()
        scheduler.add_job(restart_space, "interval", seconds=1800)
        scheduler.start()
    demo.queue(default_concurrency_limit=40).launch(ssr_mode=False)