main.py

"""
Main module for the WhisperKit Evaluation Dashboard.
This module sets up and runs the Gradio interface for the WhisperKit Evaluation Dashboard,
allowing users to explore and compare speech recognition model performance across different
devices, operating systems, and datasets.
"""

import json
import os
import re
from math import ceil, floor

import gradio as gr
import pandas as pd
from argmax_gradio_components import RangeSlider
from dotenv import load_dotenv
from huggingface_hub import login

# Import custom constants and utility functions
from constants import (
    BANNER_TEXT,
    CITATION_BUTTON_LABEL,
    CITATION_BUTTON_TEXT,
    COL_NAMES,
    HEADER,
    METHODOLOGY_TEXT,
    PERFORMANCE_TEXT,
)
from utils import (
    add_datasets_to_performance_columns,
    calculate_quality_parity,
    create_initial_performance_column_dict,
    css,
    fields,
    get_os_name_and_version,
    make_model_name_clickable_link,
    plot_metric,
    read_json_line_by_line,
)

# Load environment variables
load_dotenv()

# Get the Hugging Face token from the environment variable
HF_TOKEN = os.getenv("HF_TOKEN")

# Use the token for login
login(token=HF_TOKEN, add_to_git_credential=True)

# Define repository and directory information
repo_id = "argmaxinc/whisperkit-evals-dataset"
directory = "xcresults/benchmark_results"
local_dir = ""

# Load benchmark data from JSON files
PERFORMANCE_DATA = read_json_line_by_line("dashboard_data/performance_data.json")
with open("dashboard_data/version.json", "r") as file:
    VERSION_DATA = json.load(file)

# Load quality data (ground truth WER)
QUALITY_DATA = read_json_line_by_line("dashboard_data/quality_data.json")

SHA_TO_VERSION = {
    VERSION_DATA["releases"][i]: VERSION_DATA["versions"][i]
    for i in range(len(VERSION_DATA["versions"]))
}

# Convert JSON data to pandas DataFrames - performance only
benchmark_df = pd.json_normalize(PERFORMANCE_DATA)
releases = VERSION_DATA["releases"]

# Process timestamp data
benchmark_df["timestamp"] = pd.to_datetime(benchmark_df["timestamp"]).dt.tz_localize(
    None
)

# Use average_wer directly from performance data
benchmark_df["english_wer"] = benchmark_df["average_wer"]

sorted_performance_df = (
    benchmark_df.assign(model_len=benchmark_df["model"].str.len())
    .sort_values(
        by=["model_len", "model", "device", "os", "timestamp"],
        ascending=[True, True, True, True, False],
    )
    .drop(columns=["model_len"])
    .drop_duplicates(subset=["model", "device", "os"], keep="first")
    .reset_index(drop=True)
)

# Identify dataset-specific columns
dataset_speed_columns = [
    col for col in sorted_performance_df.columns if col.startswith("dataset_speed.")
]
dataset_toks_columns = [
    col
    for col in sorted_performance_df.columns
    if col.startswith("dataset_tokens_per_second.")
]

# Extract dataset names
PERFORMANCE_DATASETS = [col.split(".")[-1] for col in dataset_speed_columns]

# Prepare DataFrames for display
performance_df = sorted_performance_df[
    [
        "model",
        "device",
        "os",
        "english_wer",
        "qoi",
        "speed",
        "tokens_per_second",
        "timestamp",
        "commit_hash",
    ]
    + dataset_speed_columns
    + dataset_toks_columns
].copy()

# Calculate parity (difference between measured WER and ground truth WER)
performance_df["parity"] = performance_df.apply(
    lambda row: calculate_quality_parity(QUALITY_DATA, row), axis=1
)

# Rename columns for clarity
performance_df = performance_df.rename(
    lambda x: COL_NAMES[x] if x in COL_NAMES else x, axis="columns"
)

# Process dataset-specific columns
for col in dataset_speed_columns:
    dataset_name = col.split(".")[-1]
    performance_df = performance_df.rename(
        columns={
            col: f"{'Short-Form' if dataset_name == 'librispeech-10mins' else 'Long-Form'} Speed"
        }
    )

for col in dataset_toks_columns:
    dataset_name = col.split(".")[-1]
    performance_df = performance_df.rename(
        columns={
            col: f"{'Short-Form' if dataset_name == 'librispeech-10mins' else 'Long-Form'} Tok/s"
        }
    )

# Process model names for display
performance_df["model_raw"] = performance_df["Model"].copy()
performance_df["Model"] = performance_df["Model"].apply(
    lambda x: make_model_name_clickable_link(x)
)

# Extract unique devices and OS versions
initial_release_df = benchmark_df[benchmark_df["commit_hash"] == releases[-1]]
PERFORMANCE_DEVICES = initial_release_df["device"].unique().tolist()
PERFORMANCE_OS = (
    initial_release_df["os"].apply(get_os_name_and_version).unique().tolist()
)
PERFORMANCE_OS.sort()

# Create initial column dictionaries and update with dataset information
initial_performance_column_dict = create_initial_performance_column_dict()

performance_column_info = add_datasets_to_performance_columns(
    initial_performance_column_dict, PERFORMANCE_DATASETS
)

# Unpack the returned dictionaries
updated_performance_column_dict = performance_column_info["column_dict"]

PerformanceAutoEvalColumn = performance_column_info["AutoEvalColumn"]

# Define column sets for different views
PERFORMANCE_COLS = performance_column_info["COLS"]
PERFORMANCE_TYPES = performance_column_info["TYPES"]
PERFORMANCE_ALWAYS_HERE_COLS = performance_column_info["ALWAYS_HERE_COLS"]
PERFORMANCE_TOGGLE_COLS = performance_column_info["TOGGLE_COLS"]
PERFORMANCE_SELECTED_COLS = performance_column_info["SELECTED_COLS"]


def get_release_devices(release):
    """
    Get the list of devices for a specific release.

    :param release: Selected release hash
    :return: List of devices available in the release
    """
    release_df = benchmark_df[benchmark_df["commit_hash"] == release]
    return release_df["device"].unique().tolist()


def performance_filter(
    df,
    columns,
    model_query,
    exclude_models,
    devices,
    os,
    short_speed_slider,
    long_speed_slider,
    short_toks_slider,
    long_toks_slider,
    release,
):
    """
    Filters the performance DataFrame based on specified criteria.
    :param df: The DataFrame to be filtered.
    :param columns: The columns to be included in the filtered DataFrame.
    :param model_query: The query string to filter the 'Model' column.
    :param exclude_models: Models to exclude from the results.
    :param devices: The devices to filter the 'Device' column.
    :param os: The list of operating systems to filter the 'OS' column.
    :param short_speed_slider: The range of values to filter the 'Short-Form Speed' column.
    :param long_speed_slider: The range of values to filter the 'Long-Form Speed' column.
    :param short_toks_slider: The range of values to filter the 'Short-Form Tok/s' column.
    :param long_toks_slider: The range of values to filter the 'Long-Form Tok/s' column.
    :return: The filtered DataFrame.
    """
    filtered_df = df[df["commit_hash"] == release]

    # Select columns based on input and always-present columns
    filtered_df = filtered_df[
        PERFORMANCE_ALWAYS_HERE_COLS
        + [c for c in PERFORMANCE_COLS if c in df.columns and c in columns]
    ]

    # Filter models based on query
    if model_query:
        filtered_df = filtered_df[
            filtered_df["Model"].str.contains(
                "|".join(q.strip() for q in model_query.split(";")), case=False
            )
        ]

    # Exclude specified models
    if exclude_models:
        exclude_list = [m.strip() for m in exclude_models.split(";")]
        filtered_df = filtered_df[
            ~filtered_df["Model"].str.contains("|".join(exclude_list), case=False)
        ]

    # Filter by devices
    if devices:
        filtered_df = filtered_df[filtered_df["Device"].isin(devices)]
    else:
        filtered_df = pd.DataFrame(columns=filtered_df.columns)

    # Filter by operating systems
    filtered_df = (
        filtered_df[
            (
                filtered_df["OS"].str.contains(
                    "|".join(q.strip() for q in os), case=False
                )
            )
        ]
        if os
        else pd.DataFrame(columns=filtered_df.columns)
    )

    # Apply short-form and long-form speed and tokens per second filters
    min_short_speed, max_short_speed = short_speed_slider
    min_long_speed, max_long_speed = long_speed_slider
    min_short_toks, max_short_toks = short_toks_slider
    min_long_toks, max_long_toks = long_toks_slider

    df["Short-Form Speed"] = pd.to_numeric(df["Short-Form Speed"], errors="coerce")
    df["Long-Form Speed"] = pd.to_numeric(df["Long-Form Speed"], errors="coerce")
    df["Short-Form Tok/s"] = pd.to_numeric(df["Short-Form Tok/s"], errors="coerce")
    df["Long-Form Tok/s"] = pd.to_numeric(df["Long-Form Tok/s"], errors="coerce")

    if "Short-Form Speed" in filtered_df.columns:
        filtered_df = filtered_df[
            (filtered_df["Short-Form Speed"] >= min_short_speed)
            & (filtered_df["Short-Form Speed"] <= max_short_speed)
        ]
    if "Long-Form Speed" in filtered_df.columns:
        filtered_df = filtered_df[
            (filtered_df["Long-Form Speed"] >= min_long_speed)
            & (filtered_df["Long-Form Speed"] <= max_long_speed)
        ]
    if "Short-Form Tok/s" in filtered_df.columns:
        filtered_df = filtered_df[
            (filtered_df["Short-Form Tok/s"] >= min_short_toks)
            & (filtered_df["Short-Form Tok/s"] <= max_short_toks)
        ]
    if "Long-Form Tok/s" in filtered_df.columns:
        filtered_df = filtered_df[
            (filtered_df["Long-Form Tok/s"] >= min_long_toks)
            & (filtered_df["Long-Form Tok/s"] <= max_long_toks)
        ]

    return filtered_df


def update_performance_filters(release):
    """
    Updates the performance filters (devices and OS) based on the selected release.

    :param release: Selected release hash
    :return: Tuple containing updated device and OS choices
    """
    # Filter benchmark data for the selected release
    release_df = benchmark_df[benchmark_df["commit_hash"] == release]

    # Get unique devices and OS versions for this release
    release_devices = release_df["device"].unique().tolist()
    release_os = release_df["os"].apply(get_os_name_and_version).unique().tolist()
    release_os.sort()

    return (
        gr.update(choices=release_devices, value=release_devices),
        gr.update(choices=release_os, value=release_os),
    )


def update_support_table(release):
    """
    Updates the support table and its column configuration for a given release.

    :param release: Selected release hash
    :return: Tuple containing (updated DataFrame, updated column choices, updated column values)
    """
    # Load new support data
    support_data = pd.read_csv(f"dashboard_data/support_data_{release[:7]}.csv")
    support_data.set_index(support_data.columns[0], inplace=True)

    # Process model names
    support_data["Model"] = support_data["Model"].apply(lambda x: x.replace("_", "/"))
    support_data["Model"] = support_data["Model"].apply(
        lambda x: make_model_name_clickable_link(x)
    )

    # Sort by model name length
    support_data = (
        support_data.assign(model_len=support_data["Model"].str.len())
        .sort_values(
            by=["model_len"],
            ascending=[True],
        )
        .drop(columns=["model_len"])
    )

    # Get new columns (excluding 'Model')
    new_columns = support_data.columns.tolist()[1:]

    return (
        gr.update(value=support_data, datatype=["html" for _ in support_data.columns]),
        gr.update(choices=new_columns, value=new_columns),
        gr.update(value=support_data),
    )


diff_tab = gr.TabItem("Difference Checker", elem_id="diff_checker", id=2)
text_diff_elems = []

tabs = gr.Tabs(elem_id="tab-elems")

font = [
    "Zwizz Regular",  # Local font
    "IBM Plex Mono",  # Monospace font
    "ui-sans-serif",
    "system-ui",
    "sans-serif",
]


# Macos 14, 15, 26
# ios 17, 18, 26

# Define the Gradio interface
with gr.Blocks(css=css, theme=gr.themes.Base(font=font)) as demo:
    # Add header and banner to the interface
    gr.HTML(HEADER)
    gr.HTML(BANNER_TEXT, elem_classes="markdown-text")
    gr.Markdown("### Release")
    release_dropdown = gr.Dropdown(
        choices=[
            (f"{release} v{SHA_TO_VERSION[release]}", release) for release in releases
        ],
        label="Select Release",
        value=releases[-1] if releases else None,
        elem_id="release-dropdown",
        container=False,
    )

    # Create tabs for different sections of the dashboard
    with tabs.render():
        # Performance Tab
        with gr.TabItem("Benchmark", elem_id="benchmark", id=0):
            with gr.Row():
                with gr.Column(scale=1):
                    with gr.Row():
                        with gr.Column(scale=6, elem_classes="filter_models_column"):
                            filter_performance_models = gr.Textbox(
                                placeholder="🔍 Filter Model (separate multiple queries with ';')",
                                label="Filter Models",
                            )
                        with gr.Column(scale=4, elem_classes="exclude_models_column"):
                            exclude_performance_models = gr.Textbox(
                                placeholder="🔍 Exclude Model",
                                label="Exclude Model",
                            )
                    with gr.Row():
                        with gr.Accordion("See All Columns", open=False):
                            with gr.Row():
                                with gr.Column(scale=9, elem_id="performance_columns"):
                                    performance_shown_columns = gr.CheckboxGroup(
                                        choices=PERFORMANCE_TOGGLE_COLS,
                                        value=PERFORMANCE_SELECTED_COLS,
                                        label="Toggle Columns",
                                        elem_id="column-select",
                                        interactive=True,
                                    )
                                with gr.Column(
                                    scale=1,
                                    min_width=200,
                                    elem_id="performance_select_columns",
                                ):
                                    with gr.Row():
                                        select_all_button = gr.Button(
                                            "Select All",
                                            elem_id="select-all-button",
                                            interactive=True,
                                        )
                                        deselect_all_button = gr.Button(
                                            "Deselect All",
                                            elem_id="deselect-all-button",
                                            interactive=True,
                                        )

                            def select_all_columns():
                                return PERFORMANCE_TOGGLE_COLS

                            def deselect_all_columns():
                                return []

                            select_all_button.click(
                                select_all_columns,
                                inputs=[],
                                outputs=performance_shown_columns,
                            )
                            deselect_all_button.click(
                                deselect_all_columns,
                                inputs=[],
                                outputs=performance_shown_columns,
                            )

                    with gr.Row():
                        with gr.Accordion("Filter Devices", open=False):
                            with gr.Row():
                                with gr.Column(
                                    scale=9, elem_id="filter_devices_column"
                                ):
                                    performance_shown_devices = gr.CheckboxGroup(
                                        choices=get_release_devices(releases[-1]),
                                        value=get_release_devices(releases[-1]),
                                        label="Filter Devices",
                                        interactive=True,
                                    )
                                with gr.Column(
                                    scale=1,
                                    min_width=200,
                                    elem_id="filter_select_devices",
                                ):
                                    with gr.Row():
                                        select_all_devices_button = gr.Button(
                                            "Select All",
                                            elem_id="select-all-devices-button",
                                            interactive=True,
                                        )
                                        deselect_all_devices_button = gr.Button(
                                            "Deselect All",
                                            elem_id="deselect-all-devices-button",
                                            interactive=True,
                                        )

                            def select_all_devices(release):
                                """Returns all devices available in the current release"""
                                return get_release_devices(release)

                            def deselect_all_devices():
                                """Returns an empty list for deselecting all devices"""
                                return []

                            select_all_devices_button.click(
                                select_all_devices,
                                inputs=[release_dropdown],
                                outputs=performance_shown_devices,
                            )
                            deselect_all_devices_button.click(
                                deselect_all_devices,
                                inputs=[],
                                outputs=performance_shown_devices,
                            )
                    with gr.Row():
                        performance_shown_os = gr.CheckboxGroup(
                            choices=PERFORMANCE_OS,
                            value=PERFORMANCE_OS,
                            label="Filter OS",
                            interactive=True,
                        )
                with gr.Column(scale=1):
                    with gr.Accordion("See Performance Filters"):
                        with gr.Row():
                            with gr.Row():
                                min_short_speed, max_short_speed = floor(
                                    min(performance_df["Short-Form Speed"])
                                ), ceil(max(performance_df["Short-Form Speed"]))
                                short_speed_slider = RangeSlider(
                                    value=[min_short_speed, max_short_speed],
                                    minimum=min_short_speed,
                                    maximum=max_short_speed,
                                    step=0.001,
                                    label="Short-Form Speed",
                                )
                            with gr.Row():
                                min_long_speed, max_long_speed = floor(
                                    min(performance_df["Long-Form Speed"])
                                ), ceil(max(performance_df["Long-Form Speed"]))
                                long_speed_slider = RangeSlider(
                                    value=[min_long_speed, max_long_speed],
                                    minimum=min_long_speed,
                                    maximum=max_long_speed,
                                    step=0.001,
                                    label="Long-Form Speed",
                                )
                        with gr.Row():
                            with gr.Row():
                                min_short_toks, max_short_toks = floor(
                                    min(performance_df["Short-Form Tok/s"])
                                ), ceil(max(performance_df["Short-Form Tok/s"]))
                                short_toks_slider = RangeSlider(
                                    value=[min_short_toks, max_short_toks],
                                    minimum=min_short_toks,
                                    maximum=max_short_toks,
                                    step=0.001,
                                    label="Short-Form Tok/s",
                                )
                            with gr.Row():
                                min_long_toks, max_long_toks = floor(
                                    min(performance_df["Long-Form Tok/s"])
                                ), ceil(max(performance_df["Long-Form Tok/s"]))
                                long_toks_slider = RangeSlider(
                                    value=[min_long_toks, max_long_toks],
                                    minimum=min_long_toks,
                                    maximum=max_long_toks,
                                    step=0.001,
                                    label="Long-Form Tok/s",
                                )
                    with gr.Row():
                        gr.Markdown(PERFORMANCE_TEXT, elem_classes="markdown-text")
            with gr.Row():
                initial_df = performance_df[
                    performance_df["commit_hash"] == releases[-1]
                ]
                leaderboard_df = gr.components.Dataframe(
                    value=initial_df[
                        PERFORMANCE_ALWAYS_HERE_COLS + performance_shown_columns.value
                    ],
                    headers=[
                        PERFORMANCE_ALWAYS_HERE_COLS + performance_shown_columns.value
                    ],
                    datatype=[
                        c.type
                        for c in fields(PerformanceAutoEvalColumn)
                        if c.name in PERFORMANCE_COLS
                    ],
                    elem_id="leaderboard-table",
                    elem_classes="large-table",
                    interactive=False,
                )

                # Copy of the leaderboard dataframe to apply filters to
                hidden_leaderboard_df = gr.components.Dataframe(
                    value=performance_df,
                    headers=PERFORMANCE_COLS,
                    datatype=[
                        c.type
                        for c in fields(PerformanceAutoEvalColumn)
                        if c.name in PERFORMANCE_COLS
                    ],
                    visible=False,
                )

                # Inputs for the dataframe filter function
                performance_filter_inputs = [
                    hidden_leaderboard_df,
                    performance_shown_columns,
                    filter_performance_models,
                    exclude_performance_models,
                    performance_shown_devices,
                    performance_shown_os,
                    short_speed_slider,
                    long_speed_slider,
                    short_toks_slider,
                    long_toks_slider,
                    release_dropdown,
                ]

                filter_output = leaderboard_df
                filter_performance_models.change(
                    performance_filter, performance_filter_inputs, filter_output
                )
                exclude_performance_models.change(
                    performance_filter, performance_filter_inputs, filter_output
                )
                performance_shown_columns.change(
                    performance_filter, performance_filter_inputs, filter_output
                )
                performance_shown_devices.change(
                    performance_filter, performance_filter_inputs, filter_output
                )
                performance_shown_os.change(
                    performance_filter, performance_filter_inputs, filter_output
                )
                short_speed_slider.change(
                    performance_filter, performance_filter_inputs, filter_output
                )
                long_speed_slider.change(
                    performance_filter, performance_filter_inputs, filter_output
                )
                short_toks_slider.change(
                    performance_filter, performance_filter_inputs, filter_output
                )
                long_toks_slider.change(
                    performance_filter, performance_filter_inputs, filter_output
                )
                release_dropdown.change(
                    fn=update_performance_filters,
                    inputs=[release_dropdown],
                    outputs=[performance_shown_devices, performance_shown_os],
                    queue=False,
                ).then(
                    fn=performance_filter,
                    inputs=performance_filter_inputs,
                    outputs=filter_output,
                )

        # Timeline Tab
        with gr.TabItem("Timeline", elem_id="timeline", id=4):
            # Create subtabs for different metrics
            with gr.Tabs():
                with gr.TabItem("QoI", id=0):
                    with gr.Row():
                        with gr.Column(scale=6):
                            filter_qoi = gr.Textbox(
                                placeholder="🔍 Filter Model-Device-OS (separate multiple queries with ';')",
                                label="Filter",
                            )
                        with gr.Column(scale=4):
                            exclude_qoi = gr.Textbox(
                                placeholder="🔍 Exclude Model-Device-OS",
                                label="Exclude",
                            )
                    with gr.Row():
                        with gr.Column():
                            qoi_plot = gr.Plot(container=True)
                            demo.load(
                                lambda x, y, z: plot_metric(
                                    x,
                                    "qoi",
                                    "QoI",
                                    "QoI Over Time for Model-Device-OS Combinations",
                                    y,
                                    z,
                                ),
                                [
                                    gr.Dataframe(benchmark_df, visible=False),
                                    filter_qoi,
                                    exclude_qoi,
                                ],
                                qoi_plot,
                            )
                            filter_qoi.change(
                                lambda x, y, z: plot_metric(
                                    x,
                                    "qoi",
                                    "QoI",
                                    "QoI Over Time for Model-Device-OS Combinations",
                                    y,
                                    z,
                                ),
                                [
                                    gr.Dataframe(benchmark_df, visible=False),
                                    filter_qoi,
                                    exclude_qoi,
                                ],
                                qoi_plot,
                            )
                            exclude_qoi.change(
                                lambda x, y, z: plot_metric(
                                    x,
                                    "qoi",
                                    "QoI",
                                    "QoI Over Time for Model-Device-OS Combinations",
                                    y,
                                    z,
                                ),
                                [
                                    gr.Dataframe(benchmark_df, visible=False),
                                    filter_qoi,
                                    exclude_qoi,
                                ],
                                qoi_plot,
                            )

                with gr.TabItem("Average WER", id=1):
                    with gr.Row():
                        with gr.Column(scale=6):
                            filter_average_wer = gr.Textbox(
                                placeholder="🔍 Filter Model-Device-OS (separate multiple queries with ';')",
                                label="Filter",
                            )
                        with gr.Column(scale=4):
                            exclude_average_wer = gr.Textbox(
                                placeholder="🔍 Exclude Model-Device-OS",
                                label="Exclude",
                            )
                    with gr.Row():
                        with gr.Column():
                            average_wer_plot = gr.Plot(container=True)
                            demo.load(
                                lambda x, y, z: plot_metric(
                                    x,
                                    "average_wer",
                                    "Average WER",
                                    "Average WER Over Time for Model-Device-OS Combinations",
                                    y,
                                    z,
                                ),
                                [
                                    gr.Dataframe(benchmark_df, visible=False),
                                    filter_average_wer,
                                    exclude_average_wer,
                                ],
                                average_wer_plot,
                            )
                            filter_average_wer.change(
                                lambda x, y, z: plot_metric(
                                    x,
                                    "average_wer",
                                    "Average WER",
                                    "Average WER Over Time for Model-Device-OS Combinations",
                                    y,
                                    z,
                                ),
                                [
                                    gr.Dataframe(benchmark_df, visible=False),
                                    filter_average_wer,
                                    exclude_average_wer,
                                ],
                                average_wer_plot,
                            )
                            exclude_average_wer.change(
                                lambda x, y, z: plot_metric(
                                    x,
                                    "average_wer",
                                    "Average WER",
                                    "Average WER Over Time for Model-Device-OS Combinations",
                                    y,
                                    z,
                                ),
                                [
                                    gr.Dataframe(benchmark_df, visible=False),
                                    filter_average_wer,
                                    exclude_average_wer,
                                ],
                                average_wer_plot,
                            )

                with gr.TabItem("Speed", id=2):
                    with gr.Row():
                        with gr.Column(scale=6):
                            filter_speed = gr.Textbox(
                                placeholder="🔍 Filter Model-Device-OS (separate multiple queries with ';')",
                                label="Filter",
                            )
                        with gr.Column(scale=4):
                            exclude_speed = gr.Textbox(
                                placeholder="🔍 Exclude Model-Device-OS",
                                label="Exclude",
                            )
                    with gr.Row():
                        with gr.Column():
                            speed_plot = gr.Plot(container=True)
                            demo.load(
                                lambda x, y, z: plot_metric(
                                    x,
                                    "speed",
                                    "Speed",
                                    "Speed Over Time for Model-Device-OS Combinations",
                                    y,
                                    z,
                                ),
                                [
                                    gr.Dataframe(benchmark_df, visible=False),
                                    filter_speed,
                                    exclude_speed,
                                ],
                                speed_plot,
                            )
                            filter_speed.change(
                                lambda x, y, z: plot_metric(
                                    x,
                                    "speed",
                                    "Speed",
                                    "Speed Over Time for Model-Device-OS Combinations",
                                    y,
                                    z,
                                ),
                                [
                                    gr.Dataframe(benchmark_df, visible=False),
                                    filter_speed,
                                    exclude_speed,
                                ],
                                speed_plot,
                            )
                            exclude_speed.change(
                                lambda x, y, z: plot_metric(
                                    x,
                                    "speed",
                                    "Speed",
                                    "Speed Over Time for Model-Device-OS Combinations",
                                    y,
                                    z,
                                ),
                                [
                                    gr.Dataframe(benchmark_df, visible=False),
                                    filter_speed,
                                    exclude_speed,
                                ],
                                speed_plot,
                            )

                with gr.TabItem("Tok/s", id=3):
                    with gr.Row():
                        with gr.Column(scale=6):
                            filter_toks = gr.Textbox(
                                placeholder="🔍 Filter Model-Device-OS (separate multiple queries with ';')",
                                label="Filter",
                            )
                        with gr.Column(scale=4):
                            exclude_toks = gr.Textbox(
                                placeholder="🔍 Exclude Model-Device-OS",
                                label="Exclude",
                            )
                    with gr.Row():
                        with gr.Column():
                            toks_plot = gr.Plot(container=True)
                            demo.load(
                                lambda x, y, z: plot_metric(
                                    x,
                                    "tokens_per_second",
                                    "Tok/s",
                                    "Tok/s Over Time for Model-Device-OS Combinations",
                                    y,
                                    z,
                                ),
                                [
                                    gr.Dataframe(benchmark_df, visible=False),
                                    filter_toks,
                                    exclude_toks,
                                ],
                                toks_plot,
                            )
                            filter_toks.change(
                                lambda x, y, z: plot_metric(
                                    x,
                                    "tokens_per_second",
                                    "Tok/s",
                                    "Tok/s Over Time for Model-Device-OS Combinations",
                                    y,
                                    z,
                                ),
                                [
                                    gr.Dataframe(benchmark_df, visible=False),
                                    filter_toks,
                                    exclude_toks,
                                ],
                                toks_plot,
                            )
                            exclude_toks.change(
                                lambda x, y, z: plot_metric(
                                    x,
                                    "tokens_per_second",
                                    "Tok/s",
                                    "Tok/s Over Time for Model-Device-OS Combinations",
                                    y,
                                    z,
                                ),
                                [
                                    gr.Dataframe(benchmark_df, visible=False),
                                    filter_toks,
                                    exclude_toks,
                                ],
                                toks_plot,
                            )

        # Device Support Tab
        with gr.TabItem("Device Support", elem_id="device_support", id=6):
            # Add clear description of what Device Support means
            gr.Markdown(
                """
                ## Device Support
                
                This tab shows **test results for SKUs that we actually attempted to test**. It tells you whether tests passed, failed, or couldn't be completed for the devices we tried to run tests on.
                
                ### Please Note:
                **This tab only shows devices we attempted to test** - it doesn't show the full universe of available devices.
                
                **📝 For comprehensive coverage analysis**, see the **Test Coverage** tab which shows ALL available SKUs.
                """,
                elem_classes="markdown-text"
            )
            
            # Load device support data from CSV
            support_data = pd.read_csv(
                f"dashboard_data/support_data_{releases[-1][:7]}.csv"
            )
            support_data.set_index(support_data.columns[0], inplace=True)
            support_data["Model"] = support_data["Model"].apply(
                lambda x: x.replace("_", "/")
            )
            support_data["Model"] = support_data["Model"].apply(
                lambda x: make_model_name_clickable_link(x)
            )
            support_data = (
                support_data.assign(model_len=support_data["Model"].str.len())
                .sort_values(
                    by=["model_len"],
                    ascending=[True],
                )
                .drop(columns=["model_len"])
            )

            with gr.Row():
                with gr.Column(scale=1):
                    with gr.Row():
                        with gr.Column(scale=6, elem_id="filter_models_column"):
                            filter_support_models = gr.Textbox(
                                placeholder="🔍 Filter Model (separate multiple queries with ';')",
                                label="Filter Models",
                            )
                        with gr.Column(scale=4, elem_classes="exclude_models_column"):
                            exclude_support_models = gr.Textbox(
                                placeholder="🔍 Exclude Model",
                                label="Exclude Model",
                            )
                    with gr.Row():
                        with gr.Accordion("See All Columns", open=False):
                            with gr.Row():
                                with gr.Column(scale=9):
                                    support_shown_columns = gr.CheckboxGroup(
                                        choices=support_data.columns.tolist()[
                                            1:
                                        ],  # Exclude 'Model' column
                                        value=support_data.columns.tolist()[1:],
                                        label="Toggle Columns",
                                        elem_id="support-column-select",
                                        interactive=True,
                                    )
                                with gr.Column(scale=1, min_width=200):
                                    with gr.Row():
                                        select_all_support_button = gr.Button(
                                            "Select All",
                                            elem_id="select-all-support-button",
                                            interactive=True,
                                        )
                                        deselect_all_support_button = gr.Button(
                                            "Deselect All",
                                            elem_id="deselect-all-support-button",
                                            interactive=True,
                                        )

            with gr.Column():
                gr.Markdown(
                    """
                ### Legend
                - ✅ Supported: The model is supported and tested on this device.
                - ⚠️ Failed: Either the model tests failed on this device or the Speed Factor for the test is less than 1.
                - ? Not Tested: The model is supported on this device but no test information available.
                - Not Supported: The model is not supported on this device as per the [WhisperKit configuration](https://huggingface.co/argmaxinc/whisperkit-coreml/blob/main/config.json).
                """
                )

            # Display device support data in a table
            device_support_table = gr.Dataframe(
                value=support_data,
                headers=support_data.columns.tolist(),
                datatype=["html" for _ in support_data.columns],
                elem_id="device-support-table",
                elem_classes="large-table",
                interactive=False,
            )

            # Hidden dataframe to store the original data
            hidden_support_df = gr.Dataframe(value=support_data, visible=False)

            def filter_support_data(df, columns, model_query, exclude_models):
                """
                Filters the device support data based on specified criteria.

                :param df: The DataFrame to be filtered
                :param columns: Columns to include in the output
                :param model_query: Query string to filter models
                :param exclude_models: Models to exclude
                :return: Filtered DataFrame
                """
                filtered_df = df.copy()

                # Filter models based on query
                if model_query:
                    filtered_df = filtered_df[
                        filtered_df["Model"].str.contains(
                            "|".join(q.strip() for q in model_query.split(";")),
                            case=False,
                            regex=True,
                        )
                    ]

                # Exclude specified models
                if exclude_models:
                    exclude_list = [
                        re.escape(m.strip()) for m in exclude_models.split(";")
                    ]
                    filtered_df = filtered_df[
                        ~filtered_df["Model"].str.contains(
                            "|".join(exclude_list), case=False, regex=True
                        )
                    ]

                # Select columns
                selected_columns = ["Model"] + [
                    col for col in columns if col in df.columns
                ]
                filtered_df = filtered_df[selected_columns]

                return filtered_df

            def select_all_support_columns(release):
                """
                Returns all current columns from the support shown columns.

                :param release: Selected release hash
                :return: List of all available choices
                """
                # Load new support data for the current release
                support_data = pd.read_csv(
                    f"dashboard_data/support_data_{release[:7]}.csv"
                )
                support_data.set_index(support_data.columns[0], inplace=True)
                # Return all columns except 'Model'
                return [col for col in support_data.columns if col != "Model"]

            def deselect_all_support_columns():
                return []

            # Connect select all and deselect all buttons
            select_all_support_button.click(
                select_all_support_columns,
                inputs=[release_dropdown],
                outputs=support_shown_columns,
            )
            deselect_all_support_button.click(
                deselect_all_support_columns,
                inputs=[],
                outputs=support_shown_columns,
            )

            # Connect release dropdown to support data update
            release_dropdown.change(
                update_support_table,
                inputs=[release_dropdown],
                outputs=[
                    device_support_table,
                    support_shown_columns,
                    hidden_support_df,
                ],
            ).then(
                filter_support_data,
                inputs=[
                    hidden_support_df,
                    support_shown_columns,
                    filter_support_models,
                    exclude_support_models,
                ],
                outputs=device_support_table,
            )

            # Also connect the filter inputs to update the table
            for input_elem in [
                filter_support_models,
                exclude_support_models,
                support_shown_columns,
            ]:
                input_elem.change(
                    filter_support_data,
                    inputs=[
                        hidden_support_df,
                        support_shown_columns,
                        filter_support_models,
                        exclude_support_models,
                    ],
                    outputs=device_support_table,
                )

        # Test Coverage Tab
        with gr.TabItem("Test Coverage", elem_id="test_coverage", id=7):
            # Add clear description of what Test Coverage means
            gr.Markdown(
                """
                ## Test Coverage 
                
                This tab shows **ALL available SKUs** and our testing coverage across the entire device ecosystem. Uses chip-based expansion where testing one device covers all devices with the same chip.
                """,
                elem_classes="markdown-text"
            )

            def load_coverage_data(release):
                """Load test coverage data for a specific release."""
                try:
                    with open(f"dashboard_data/test_coverage_{release}.json", "r") as f:
                        return json.load(f)
                except FileNotFoundError:
                    return {
                        "commit_hash": release,
                        "total_devices": 0,
                        "tested_devices": 0,
                        "skipped_devices": 0,
                        "coverage_percentage": 0.0,
                        "tested_device_list": [],
                        "skipped_device_list": [],
                        "tested_os_versions": [],
                        "has_target_os_coverage": False,
                        "covered_target_versions": [],
                        "missing_target_versions": [],
                    }

            def format_coverage_devices(device_list):
                """Convert device list to DataFrame format."""
                if not device_list:
                    return pd.DataFrame(columns=["Device"])

                df = pd.DataFrame({"Device": device_list})
                return df.sort_values(["Device"])

            def update_coverage_data(release):
                """Update coverage data when release changes."""
                coverage_data = load_coverage_data(release)

                # Format tested and skipped devices
                tested_df = format_coverage_devices(coverage_data["tested_device_list"])
                skipped_df = format_coverage_devices(
                    coverage_data["skipped_device_list"]
                )

                # Check target OS coverage
                target_os_status = ""
                covered_versions = coverage_data.get("covered_target_versions", [])
                missing_versions = coverage_data.get("missing_target_versions", [])
                
                if covered_versions or missing_versions:
                    target_os_status = "\n- **Target OS Coverage**:\n"
                    if covered_versions:
                        unique_versions = sorted(set(covered_versions))
                        target_os_status += f"  - ✅ **Tested**: {', '.join(unique_versions)}\n"
                    if missing_versions:
                        target_os_status += f"  - ❌ **Missing**: {', '.join(missing_versions)}"

                # Create coverage summary
                coverage_summary = f"""## Test Coverage Summary for Release {release} (v{SHA_TO_VERSION.get(release, 'Unknown')})

- **Total Devices**: {coverage_data['total_devices']}
- **Tested Devices**: {coverage_data['tested_devices']}
- **Skipped Devices**: {coverage_data['skipped_devices']}
- **Coverage Percentage**: {coverage_data['coverage_percentage']:.1f}%
{target_os_status}"""

                return (
                    gr.update(value=coverage_summary),
                    gr.update(value=tested_df),
                    gr.update(value=skipped_df),
                    tested_df,
                    skipped_df,
                )

            def filter_coverage_devices(df, device_query, exclude_devices):
                """Filter coverage devices based on device queries."""
                if df is None or df.empty:
                    return df

                filtered_df = df.copy()

                # Filter devices based on query
                if device_query:
                    filtered_df = filtered_df[
                        filtered_df["Device"].str.contains(
                            "|".join(q.strip() for q in device_query.split(";")),
                            case=False,
                            regex=True,
                        )
                    ]

                # Exclude specified devices
                if exclude_devices:
                    exclude_list = [
                        re.escape(d.strip()) for d in exclude_devices.split(";")
                    ]
                    filtered_df = filtered_df[
                        ~filtered_df["Device"].str.contains(
                            "|".join(exclude_list), case=False, regex=True
                        )
                    ]

                return filtered_df

            # Load initial coverage data
            initial_coverage = load_coverage_data(releases[-1])
            initial_tested_df = format_coverage_devices(
                initial_coverage["tested_device_list"]
            )
            initial_skipped_df = format_coverage_devices(
                initial_coverage["skipped_device_list"]
            )

            # Generate initial target OS status
            initial_target_os_status = ""
            covered_versions = initial_coverage.get("covered_target_versions", [])
            missing_versions = initial_coverage.get("missing_target_versions", [])
            
            if covered_versions or missing_versions:
                initial_target_os_status = "\n- **Target OS Coverage**:\n"
                if covered_versions:
                    unique_versions = sorted(set(covered_versions))
                    initial_target_os_status += f"  - ✅ **Tested**: {', '.join(unique_versions)}\n"
                if missing_versions:
                    initial_target_os_status += f"  - ❌ **Missing**: {', '.join(missing_versions)}"

            # Create initial coverage summary content
            initial_summary_content = f"""## Test Coverage Summary for Release {releases[-1]} (v{SHA_TO_VERSION.get(releases[-1], 'Unknown')})

- **Total Devices**: {initial_coverage['total_devices']}
- **Tested Devices**: {initial_coverage['tested_devices']}
- **Skipped Devices**: {initial_coverage['skipped_devices']}
- **Coverage Percentage**: {initial_coverage['coverage_percentage']:.1f}%
{initial_target_os_status}"""
            
            # Coverage summary
            coverage_summary_text = gr.Markdown(
                value=initial_summary_content,
                elem_classes="markdown-text"
            )

            with gr.Row():
                with gr.Column(scale=1):
                    with gr.Row():
                        with gr.Column(scale=6):
                            filter_coverage_devices_input = gr.Textbox(
                                placeholder="🔍 Filter Device (separate multiple queries with ';')",
                                label="Filter Devices",
                            )
                        with gr.Column(scale=4):
                            exclude_coverage_devices_input = gr.Textbox(
                                placeholder="🔍 Exclude Device",
                                label="Exclude Device",
                            )

            # Create tabs for tested vs skipped devices
            with gr.Tabs():
                with gr.TabItem("Tested Devices", id=0):
                    tested_devices_table = gr.Dataframe(
                        value=initial_tested_df,
                        headers=["Device"],
                        datatype=["str"],
                        elem_id="tested-devices-table",
                        elem_classes="large-table",
                        interactive=False,
                    )

                with gr.TabItem("Skipped Devices", id=1):
                    skipped_devices_table = gr.Dataframe(
                        value=initial_skipped_df,
                        headers=["Device"],
                        datatype=["str"],
                        elem_id="skipped-devices-table",
                        elem_classes="large-table",
                        interactive=False,
                    )

            # Hidden dataframes for filtering
            hidden_tested_df = gr.Dataframe(value=initial_tested_df, visible=False)
            hidden_skipped_df = gr.Dataframe(value=initial_skipped_df, visible=False)

            # Connect release dropdown to coverage data update
            release_dropdown.change(
                update_coverage_data,
                inputs=[release_dropdown],
                outputs=[
                    coverage_summary_text,
                    tested_devices_table,
                    skipped_devices_table,
                    hidden_tested_df,
                    hidden_skipped_df,
                ],
                queue=False,
            )

            # Connect filter inputs to update both tables
            for input_elem in [
                filter_coverage_devices_input,
                exclude_coverage_devices_input,
            ]:
                input_elem.change(
                    lambda tested_df, skipped_df, device_query, exclude_devices: (
                        filter_coverage_devices(
                            tested_df, device_query, exclude_devices
                        ),
                        filter_coverage_devices(
                            skipped_df, device_query, exclude_devices
                        ),
                    ),
                    inputs=[
                        hidden_tested_df,
                        hidden_skipped_df,
                        filter_coverage_devices_input,
                        exclude_coverage_devices_input,
                    ],
                    outputs=[tested_devices_table, skipped_devices_table],
                )

        # Methodology Tab
        with gr.TabItem("Methodology", elem_id="methodology", id=8):
            gr.Markdown(METHODOLOGY_TEXT, elem_id="methodology-text")

    # Citation section
    with gr.Accordion("📙 Citation", open=False):
        citation_button = gr.Textbox(
            value=CITATION_BUTTON_TEXT,
            label=CITATION_BUTTON_LABEL,
            lines=7,
            elem_id="citation-button",
            show_copy_button=True,
        )

# Launch the Gradio interface
demo.launch(debug=True, share=True, ssr_mode=False)