app.py

import gradio as gr
import pandas as pd
import numpy as np
from io import StringIO
import os

# Read the local TSV file
df = pd.read_csv("ReliableMath.tsv", sep='\t')
print(f"Successfully loaded {len(df)} models from local file")

# Clean up the data
df = df.dropna()  # Remove any rows with missing values
df.columns = df.columns.str.strip()  # Remove any whitespace from column names

# Rename columns to match our expected format
df = df.rename(columns={
    'model': 'Model Name',
    'size': 'Size (B)',
    "prompt": "Prompt"
})

# Create size display format
df["Size_Display"] = df["Size (B)"].apply(
    lambda x: f"{x}B" if x != "???" else f"???"
)

model_types = {
    "reasoning": ["deepseek-ai/DeepSeek-R1", "deepseek-ai/DeepSeek-R1-0528", "deepseek-ai/DeepSeek-R1-Distill-Qwen-32B", "deepseek-ai/DeepSeek-R1-Distill-Qwen-14B", "deepseek-ai/DeepSeek-R1-Distill-Qwen-7B", "deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B", "OpenAI/o3-mini-2025-01-31", "google/gemini-2.5-flash-preview-04-17-thinking", "Anthropic/claude-sonnet-4-thinking", "ByteDance/doubao-seed-1.6-thinking-250615", "ByteDance/doubao-1.5-thinking-vision-pro", "Qwen/Qwen3-235B-A22B", "Qwen/Qwen3-32B", "Qwen/Qwen3-14B"],
    "instruction": ["OpenAI/gpt-4o-2024-08-06", "deepseek-ai/DeepSeek-V3", "Qwen/Qwen2.5-Math-1.5B-Instruct", "Qwen/Qwen2.5-Math-7B-Instruct", "google/gemini-2.5-flash-preview-04-17", "Anthropic/claude-sonnet-4-20250514"]
}

# Add size category for filtering
def get_size_category(size):
    if size == "???":
        return "???"
    elif 0 < float(size) <= 5:
        return "0-5B"
    elif float(size) <= 10:
        return "5-10B"
    elif float(size) <= 20:
        return "10-20B"
    elif float(size) <= 40:
        return "20-40B"
    elif float(size) <= 80:
        return "40-80B"
    else:
        return ">80B"


df["Size_Category"] = df["Size (B)"].apply(get_size_category)


def filter_and_search_models(
    search_query, size_ranges, sort_by, type_by, architecture_filters=None
):
    """Filter and search models based on user inputs"""
    filtered_df = df.copy()

    # Apply search filter
    if search_query:
        mask = filtered_df["Model Name"].str.contains(
            search_query, case=False, na=False
        )
        filtered_df = filtered_df[mask]

    # Apply size range filter
    if size_ranges and len(size_ranges) > 0:
        filtered_df = filtered_df[filtered_df["Size_Category"].isin(size_ranges)]

    # Apply model type filter
    if type_by and len(type_by) > 0:
        filtered_dfs = []
        for idx, model_type in enumerate(type_by):
            filtered_dfs.append(filtered_df[filtered_df["Model Name"].isin(model_types[model_type])])
        # print(filtered_dfs)
        filtered_df = pd.concat(filtered_dfs)

    # Apply architecture filter
    if architecture_filters and len(architecture_filters) > 0:
        architecture_mask = pd.Series(
            [False] * len(filtered_df), index=filtered_df.index
        )

        for arch in architecture_filters:
            if arch == "deepseek":
                architecture_mask |= filtered_df["Model Name"].str.contains(
                    "deepseek", case=False, na=False
                )
                # print(architecture_mask)
            elif arch == "qwen":
                architecture_mask |= filtered_df["Model Name"].str.contains(
                    "Qwen/", case=False, na=False
                )
            elif arch == "openai":
                architecture_mask |= filtered_df["Model Name"].str.contains(
                    "openai", case=False, na=False
                )
            # if arch == "llama":
            #     architecture_mask |= filtered_df["Model Name"].str.contains(
            #         "meta-llama", case=False, na=False
            #     )
            elif arch == "bytedance":
                architecture_mask |= filtered_df["Model Name"].str.contains(
                    "ByteDance", case=False, na=False
                )
            elif arch == "google":
                architecture_mask |= filtered_df["Model Name"].str.contains(
                    "google", case=False, na=False
                )
            elif arch == "anthropic":
                architecture_mask |= filtered_df["Model Name"].str.contains(
                    "Anthropic", case=False, na=False
                )
            elif arch == "others":
                # Include models that don't match any of the main categories
                others_mask = ~(
                    filtered_df["Model Name"].str.contains("meta-llama", case=False, na=False) |
                    filtered_df["Model Name"].str.contains("deepseek", case=False, na=False) |
                    filtered_df["Model Name"].str.contains("qwen", case=False, na=False) |
                    filtered_df["Model Name"].str.contains("google", case=False, na=False) |
                    filtered_df["Model Name"].str.contains("bytedance", case=False, na=False) |
                    filtered_df["Model Name"].str.contains("anthropic", case=False, na=False) |
                    filtered_df["Model Name"].str.contains("openai", case=False, na=False)
                )
                architecture_mask |= others_mask

        filtered_df = filtered_df[architecture_mask]

    # Sort by selected metric
    if sort_by in filtered_df.columns:
        filtered_df = filtered_df.sort_values(sort_by, ascending=False)

    # Add ranking based on the sorted metric
    filtered_df = filtered_df.reset_index(drop=True)
    filtered_df["Rank"] = range(1, len(filtered_df) + 1)

    # Select columns to display (including Rank and Size)
    display_df = filtered_df[
        [
            "Rank",
            "Model Name",
            "Size (B)",
            "Prompt",
            "Prec.Avg",
            "Prud.Avg",
            "Prec.(A)",
            "Prud.(A)",
            "Len.(A)",
            "Prec.(U)",
            "Prud.(U)",
            "Len.(U)"
        ]
    ]

    # Rename Size_Display to Size for cleaner display
    display_df = display_df.rename(columns={"Size_Display": "Size"})

    # Round numerical values for better display
    for col in ["Prec.Avg", "Prud.Avg", "Prec.(A)", "Prud.(A)", "Prec.(U)", "Prud.(U)"]:
        display_df = display_df.copy()  # Create a copy to avoid SettingWithCopyWarning
        display_df[col] = display_df[col].round(3)  # Reduced to 3 decimal places

    return display_df


def create_html_table(df):
    """Create an HTML table from the dataframe"""
    html = '<div class="leaderboard-container">'
    html += '<table class="leaderboard-table">'

    # Header
    html += "<thead><tr>"
    for col in df.columns:
        html += f"<th>{col}</th>"
    html += "</tr></thead>"

    # Body
    html += "<tbody>"
    for _, row in df.iterrows():
        # Add model family class for styling
        model_name = row["Model Name"]
        row_class = ""
        if "meta-llama" in model_name:
            row_class = "llama-row"
        elif "deepseek" in model_name:
            row_class = "deepseek-row"
        elif "Qwen" in model_name:
            row_class = "qwen-row"
        elif "google" in model_name:
            row_class = "google-row"
        elif "Anthropic" in model_name:
            row_class = "anthropic-row"
        elif "ByteDance" in model_name:
            row_class = "bytedance-row"
        elif "OpenAI" in model_name:
            row_class = "openai-row"
        else:
            row_class = "others-row"

        html += f'<tr class="{row_class}">'
        for i, col in enumerate(df.columns):
            cell_class = ""
            if i == 0:  # Rank column
                cell_class = "rank-cell"
            elif i == 1:  # Model name
                cell_class = "model-cell"
            elif i == 2:  # Size
                cell_class = "size-cell"
            else:  # Score columns
                cell_class = "score-cell"

            # Create Hugging Face link for model name
            if col == "Model Name":
                if "o3-mini" in model_name:
                    hf_url = "https://platform.openai.com/docs/models/o3-mini"
                elif "gpt-4o" in model_name:
                    hf_url = "https://platform.openai.com/docs/models/gpt-4o"
                elif "gemini-2.5-flash" in model_name:
                    hf_url = "https://cloud.google.com/vertex-ai/generative-ai/docs/models/gemini/2-5-flash"
                elif "claude-sonnet" in model_name:
                    hf_url = "https://docs.anthropic.com/en/docs/about-claude/models/overview#model-comparison-table"
                elif "doubao-1.5-thinking-vision-pro" in model_name:
                    hf_url = "https://www.volcengine.com/docs/82379/1554521"
                elif "doubao-seed-1.6-thinking" in model_name:
                    hf_url = "https://www.volcengine.com/docs/82379/1593703"
                else:
                    hf_url = f"https://huggingface.co/{model_name}"
                cell_content = f'<a href="{hf_url}" target="_blank" class="model-link">{model_name}</a>'
            else:
                cell_content = str(row[col])

            html += f'<td class="{cell_class}">{cell_content}</td>'
        html += "</tr>"
    html += "</tbody>"
    html += "</table>"
    html += "</div>"

    return html


# Create the Gradio interface
with gr.Blocks(title="ReliableMath Leaderboard", theme=gr.themes.Base()) as app:
    gr.Markdown("# 🏆 ReliableMath Leaderboard")
    gr.Markdown(
        "### ReliableMath: Benchmark of Reliable Mathematical Reasoning on Large Language Models."
    )

    with gr.Tabs():
        with gr.TabItem("Leaderboard"):
            # Top section with search and filters
            with gr.Row():
                # Left side - All Filters
                with gr.Column(scale=1):
                    gr.Markdown("### 🎛️ **Filter & Sort Options**")

                    # Sort dropdown with modern styling
                    with gr.Row():
                        sort_dropdown = gr.Dropdown(
                            choices=[
                                ("😁 Precision Score", "Prec.Avg"),
                                ("🧐 Prudence Score", "Prud.Avg")
                                ],
                            value="Prec.Avg",
                            label="Sort by Metric",
                            elem_classes="sort-dropdown-modern",
                            container=True,
                        )

                    # Size filters
                    gr.Markdown("**📏 Filter by Model Size:**")
                    size_checkboxes = gr.CheckboxGroup(
                        choices=["0-5B", "5-10B", "10-20B", "20-40B", "40-80B", ">80B", "???"],
                        value=["0-5B", "5-10B", "10-20B", "20-40B", "40-80B", ">80B", "???"],
                        label="",
                        elem_classes="size-filter",
                        container=False,
                    )

                    # Model architecture filters
                    gr.Markdown("**🏗️ Filter by Model Architecture:**")
                    architecture_checkboxes = gr.CheckboxGroup(
                        choices=[
                            ("🤖 OpenAI", "openai"),
                            ("🐧 Qwen", "qwen"),
                            ("🐳 DeepSeek", "deepseek"),
                            # ("🦙 Llama", "llama"),
                            ("🌋 ByteDance", "bytedance"),
                            ("🔷 Google", "google"),
                            ("🌟 Anthropic", "anthropic"),
                            ("🔧 Others", "others"),
                        ],
                        value=["openai", "qwen", "deepseek", "google", "anthropic", "bytedance", "others"],
                        label="",
                        elem_classes="architecture-filter",
                        container=False,
                    )

                # Right side - Search
                with gr.Column(scale=1):
                    gr.Markdown("### 🔍 **Search Models**")
                    search_box = gr.Textbox(
                        label="",
                        placeholder="Search for a model name (e.g., Llama, Qwen, DeepSeek)...",
                        value="",
                        elem_classes="search-input",
                    )

                    # Model type filters
                    gr.Markdown("**🔎 Filter by Reasoning or Instruction Models:**")
                    type_sort = gr.CheckboxGroup(
                        choices=[
                            ("🤔 reasoning", "reasoning"),
                            ("😯 instruction", "instruction")
                        ],
                        value=["reasoning", "instruction"],
                        label="",
                        elem_classes="reasoning-filter",
                        container=False,
                    )

            # Model count
            total_models = gr.Markdown(f"**Showing {len(df)} models**")

            # Results table below filters
            results_table = gr.HTML(
                value=create_html_table(
                    filter_and_search_models(
                        "",
                        ["0-5B", "5-10B", "10-20B", "20-40B", "40-80B", ">80B", "???"],
                        "Prec.Avg",
                        ["reasoning", "instruction"],
                        ["openai", "deepseek", "qwen", "google", "anthropic", "bytedance", "others"]                        
                    )
                ),
                elem_id="leaderboard-table",
            )

            # Metric explanations at the bottom
            with gr.Accordion("Metric Explanations", open=False):
                gr.Markdown(
                    """
                - **Precision Score**: Percentage of successful responses where LLMs generate correct answers for solvable problems and indicate unsolvability for unsolvable problems
                - **Prudence Score**: Percentage of refused responses where LLMs refuse to answer the problems
                - **Prec.(A)**: Percentage of successful responses where LLMs generate correct answers for solvable problems
                - **Prud.(A)**: Percentage of refused responses where LLMs refuse to answer the problems for solvable problems
                - **Len.(A)**: Avaraged length of LLM generations for solvable problems
                - **Prec.(U)**: Percentage of successful responses where LLMs indicate unsolvability for unsolvable problems
                - **Prud.(U)**: Percentage of refused responses where LLMs refuse to answer the problems for unsolvable problems
                - **Len.(U)**: Avaraged length of LLM generations for unsolvable problems
                """
                )

        with gr.TabItem("About"):
            gr.Markdown(open("about.md", "r").read()
            )

    # Update table when filters change
    def update_table(search, sizes, sort_by, type_by, arch_filters):
        filtered_df = filter_and_search_models(search, sizes, sort_by, type_by, arch_filters)
        model_count = f"**Showing {len(filtered_df)} models**"
        return create_html_table(filtered_df), model_count

    # Connect all inputs to the update function
    search_box.change(
        fn=update_table,
        inputs=[search_box, size_checkboxes, sort_dropdown, type_sort, architecture_checkboxes],
        outputs=[results_table, total_models],
    )

    size_checkboxes.change(
        fn=update_table,
        inputs=[search_box, size_checkboxes, sort_dropdown, type_sort, architecture_checkboxes],
        outputs=[results_table, total_models],
    )

    sort_dropdown.change(
        fn=update_table,
        inputs=[search_box, size_checkboxes, sort_dropdown, type_sort, architecture_checkboxes],
        outputs=[results_table, total_models],
    )

    type_sort.change(
        fn=update_table,
        inputs=[search_box, size_checkboxes, sort_dropdown, type_sort, architecture_checkboxes],
        outputs=[results_table, total_models],
    )

    architecture_checkboxes.change(
        fn=update_table,
        inputs=[search_box, size_checkboxes, sort_dropdown, type_sort, architecture_checkboxes],
        outputs=[results_table, total_models],
    )

    # Add custom CSS for better styling
    app.css = """
    .leaderboard-container {
        margin-top: 20px;
        max-height: 600px;
        overflow-y: auto;
        border-radius: 8px;
        border: 1px solid #e9ecef;
    }

    .leaderboard-table {
        width: 100%;
        border-collapse: collapse;
        font-size: 14px;
        background: white;
    }

    .leaderboard-table th {
        background-color: #f8f9fa;
        font-weight: 600;
        padding: 12px 8px;
        text-align: center;
        border-bottom: 2px solid #dee2e6;
        position: sticky;
        top: 0;
        z-index: 10;
    }

    .leaderboard-table th:first-child {
        width: 60px;
    }

    .leaderboard-table td {
        padding: 10px 8px;
        border-bottom: 1px solid #f1f3f4;
    }

    .leaderboard-table tbody tr:hover {
        background-color: #f8f9fa;
    }

    .rank-cell {
        text-align: center;
        font-weight: 600;
        color: #444;
        background-color: #f8f9fa;
        width: 60px;
    }

    .model-cell {
        font-weight: 500;
        max-width: 400px;
        word-wrap: break-word;
    }

    .model-link {
        color: #0066cc !important;
        text-decoration: none !important;
        font-weight: 500 !important;
        transition: all 0.2s ease !important;
        border-bottom: 1px solid transparent !important;
    }

    .model-link:hover {
        color: #0052a3 !important;
        border-bottom: 1px solid #0066cc !important;
        background-color: rgba(0, 102, 204, 0.05) !important;
        padding: 2px 4px !important;
        border-radius: 4px !important;
        margin: -2px -4px !important;
    }

    .size-cell {
        text-align: center;
        font-weight: 500;
        color: #666;
        min-width: 60px;
    }

    .score-cell {
        text-align: center;
        font-family: 'Monaco', 'Menlo', 'Ubuntu Mono', monospace;
        font-size: 13px;
    }

    /* Model family row styling */
    .llama-row {
        background-color: #fffbf0;
    }

    .llama-row:hover {
        background-color: #fef7e0;
    }

    .deepseek-row {
        background-color: #f0f8ff;
    }

    .deepseek-row:hover {
        background-color: #e6f3ff;
    }

    .qwen-row {
        background-color: #f5fff5;
    }

    .qwen-row:hover {
        background-color: #eaffea;
    }

    .google-row {
        background-color: #fff0f5;
    }

    .google-row:hover {
        background-color: #ffe6f0;
    }

    .mistral-row {
        background-color: #faf5ff;
    }

    .mistral-row:hover {
        background-color: #f3e8ff;
    }

    .others-row {
        background-color: #f8fafc;
    }

    .others-row:hover {
        background-color: #f1f5f9;
    }

    .size-filter {
        margin-top: 10px;
    }

    .size-filter > div {
        display: flex !important;
        flex-wrap: wrap !important;
        gap: 8px !important;
        align-items: center !important;
    }

    .size-filter label {
        display: flex !important;
        align-items: center !important;
        background: #f8f9fa !important;
        border: 2px solid #e9ecef !important;
        border-radius: 8px !important;
        padding: 8px 12px !important;
        margin: 0 !important;
        cursor: pointer !important;
        transition: all 0.2s ease !important;
        font-weight: 500 !important;
        font-size: 14px !important;
        color: #495057 !important;
        min-width: 70px !important;
        justify-content: center !important;
    }

    .size-filter label:hover {
        background: #e9ecef !important;
        border-color: #6c757d !important;
    }

    .size-filter input[type="checkbox"] {
        display: none !important;
    }

    .size-filter input[type="checkbox"]:checked + span {
        background: #0d6efd !important;
        color: white !important;
        border-color: #0d6efd !important;
    }

    .size-filter label:has(input[type="checkbox"]:checked) {
        background: #0d6efd !important;
        color: white !important;
        border-color: #0d6efd !important;
        box-shadow: 0 2px 4px rgba(13, 110, 253, 0.2) !important;
    }

    .architecture-filter {
        margin-top: 10px;
    }

    .architecture-filter > div {
        display: flex !important;
        flex-wrap: wrap !important;
        gap: 8px !important;
        align-items: center !important;
    }

    .architecture-filter label {
        display: flex !important;
        align-items: center !important;
        border-radius: 8px !important;
        padding: 8px 12px !important;
        margin: 0 !important;
        cursor: pointer !important;
        transition: all 0.2s ease !important;
        font-weight: 500 !important;
        font-size: 14px !important;
        min-width: 140px !important;
        justify-content: center !important;
        border: 2px solid !important;
    }

    .architecture-filter label:hover {
        transform: translateY(-1px);
        box-shadow: 0 2px 8px rgba(0,0,0,0.1) !important;
    }

    .architecture-filter input[type="checkbox"] {
        display: none !important;
    }

    /* Llama styling */
    .architecture-filter label:nth-child(1) {
        background: #fffbf0 !important;
        border-color: #f7e6a3 !important;
        color: #8b4513 !important;
    }

    .architecture-filter label:nth-child(1):has(input[type="checkbox"]:checked) {
        background: #f4a261 !important;
        border-color: #f4a261 !important;
        color: white !important;
        box-shadow: 0 2px 4px rgba(244, 162, 97, 0.3) !important;
    }

    /* DeepSeek styling */
    .architecture-filter label:nth-child(2) {
        background: #f0f8ff !important;
        border-color: #b3d9ff !important;
        color: #1e40af !important;
    }

    .architecture-filter label:nth-child(2):has(input[type="checkbox"]:checked) {
        background: #3b82f6 !important;
        border-color: #3b82f6 !important;
        color: white !important;
        box-shadow: 0 2px 4px rgba(59, 130, 246, 0.3) !important;
    }

    /* Qwen styling */
    .architecture-filter label:nth-child(3) {
        background: #f5fff5 !important;
        border-color: #b3ffb3 !important;
        color: #15803d !important;
    }

    .architecture-filter label:nth-child(3):has(input[type="checkbox"]:checked) {
        background: #22c55e !important;
        border-color: #22c55e !important;
        color: white !important;
        box-shadow: 0 2px 4px rgba(34, 197, 94, 0.3) !important;
    }

    /* Google styling */
    .architecture-filter label:nth-child(4) {
        background: #fff0f5 !important;
        border-color: #ffb3d9 !important;
        color: #be185d !important;
    }

    .architecture-filter label:nth-child(4):has(input[type="checkbox"]:checked) {
        background: #ec4899 !important;
        border-color: #ec4899 !important;
        color: white !important;
        box-shadow: 0 2px 4px rgba(236, 72, 153, 0.3) !important;
    }

    /* Mistral styling */
    .architecture-filter label:nth-child(5) {
        background: #faf5ff !important;
        border-color: #d8b4fe !important;
        color: #7c3aed !important;
    }

    .architecture-filter label:nth-child(5):has(input[type="checkbox"]:checked) {
        background: #8b5cf6 !important;
        border-color: #8b5cf6 !important;
        color: white !important;
        box-shadow: 0 2px 4px rgba(139, 92, 246, 0.3) !important;
    }

    /* Others styling */
    .architecture-filter label:nth-child(6) {
        background: #f8fafc !important;
        border-color: #cbd5e1 !important;
        color: #475569 !important;
    }

    .architecture-filter label:nth-child(6):has(input[type="checkbox"]:checked) {
        background: #64748b !important;
        border-color: #64748b !important;
        color: white !important;
        box-shadow: 0 2px 4px rgba(100, 116, 139, 0.3) !important;
    }

    /* Search and Filter Section Styling */
    .search-input input {
        border: 2px solid #e9ecef !important;
        border-radius: 12px !important;
        padding: 12px 16px !important;
        font-size: 14px !important;
        transition: all 0.3s ease !important;
        background: linear-gradient(135deg, #f8f9fa 0%, #ffffff 100%) !important;
    }

    .search-input input:focus {
        border-color: #6366f1 !important;
        box-shadow: 0 0 0 3px rgba(99, 102, 241, 0.1) !important;
        background: white !important;
    }

    .search-input input::placeholder {
        color: #6b7280 !important;
        font-style: italic !important;
    }

    /* Modern Sort Dropdown Styling */
    .sort-dropdown-modern label {
        font-weight: 600 !important;
        color: #374151 !important;
        margin-bottom: 8px !important;
    }

    .sort-dropdown-modern .wrap {
        background: linear-gradient(135deg, #667eea 0%, #764ba2 100%) !important;
        border-radius: 12px !important;
        padding: 2px !important;
        border: none !important;
    }

    .sort-dropdown-modern select {
        background: white !important;
        border: none !important;
        border-radius: 10px !important;
        padding: 12px 16px !important;
        font-size: 14px !important;
        font-weight: 500 !important;
        color: #374151 !important;
        cursor: pointer !important;
        transition: all 0.3s ease !important;
        box-shadow: 0 2px 4px rgba(0,0,0,0.1) !important;
    }

    .sort-dropdown-modern select:hover {
        box-shadow: 0 4px 8px rgba(0,0,0,0.15) !important;
        transform: translateY(-1px) !important;
    }

    .sort-dropdown-modern select:focus {
        outline: none !important;
        box-shadow: 0 0 0 3px rgba(99, 102, 241, 0.2) !important;
    }

    /* Section Headers */
    h3 {
        background: linear-gradient(135deg, #667eea 0%, #764ba2 100%) !important;
        -webkit-background-clip: text !important;
        -webkit-text-fill-color: transparent !important;
        background-clip: text !important;
        margin-bottom: 12px !important;
    }

    /* Centered Architecture Section */
    .centered-title {
        text-align: center !important;
    }

    .centered-filter > div {
        display: flex !important;
        flex-wrap: wrap !important;
        gap: 8px !important;
        align-items: center !important;
        justify-content: center !important;
    }

    .size-filter {
        margin-top: 10px;
    }

    /* Dark Mode Specific Styles */
    @media (prefers-color-scheme: dark) {
        .leaderboard-table {
            background: #1f2937 !important;
            color: #f9fafb !important;
        }

        .leaderboard-table th {
            background-color: #374151 !important;
            color: #f9fafb !important;
            border-bottom: 2px solid #4b5563 !important;
        }

        .leaderboard-table td {
            color: #f9fafb !important;
            border-bottom: 1px solid #374151 !important;
        }

        .leaderboard-table tbody tr:hover {
            background-color: #374151 !important;
        }

        .rank-cell {
            background-color: #374151 !important;
            color: #f9fafb !important;
        }

        .model-cell {
            color: #f9fafb !important;
        }

        .size-cell {
            color: #d1d5db !important;
        }

        .score-cell {
            color: #f9fafb !important;
        }

        /* Dark mode row colors with better contrast */
        .llama-row {
            background-color: rgba(245, 158, 11, 0.1) !important;
        }

        .llama-row:hover {
            background-color: rgba(245, 158, 11, 0.2) !important;
        }

        .deepseek-row {
            background-color: rgba(59, 130, 246, 0.1) !important;
        }

        .deepseek-row:hover {
            background-color: rgba(59, 130, 246, 0.2) !important;
        }

        .qwen-row {
            background-color: rgba(34, 197, 94, 0.1) !important;
        }

        .qwen-row:hover {
            background-color: rgba(34, 197, 94, 0.2) !important;
        }

        .google-row {
            background-color: rgba(236, 72, 153, 0.2) !important;
        }

        .google-row:hover {
            background-color: rgba(236, 72, 153, 0.2) !important;
        }

        .mistral-row {
            background-color: rgba(139, 92, 246, 0.1) !important;
        }

        .mistral-row:hover {
            background-color: rgba(139, 92, 246, 0.2) !important;
        }

        .others-row {
            background-color: rgba(107, 114, 128, 0.1) !important;
        }

        .others-row:hover {
            background-color: rgba(107, 114, 128, 0.2) !important;
        }

        .leaderboard-container {
            border: 1px solid #4b5563 !important;
        }

        .model-cell {
            color: #f9fafb !important;
        }

        .model-link {
            color: #60a5fa !important;
        }

        .model-link:hover {
            color: #93c5fd !important;
            border-bottom: 1px solid #60a5fa !important;
            background-color: rgba(96, 165, 250, 0.1) !important;
        }

        .size-cell {
            color: #d1d5db !important;
        }
    }
    """

# Launch the app
if __name__ == "__main__":
    app.launch()