app.py

from __future__ import annotations

import base64
import json
from functools import cmp_to_key
from html import escape
from pathlib import Path
from typing import Any

import gradio as gr


HOME_VIEW = "HOME"
TASK_ORDER = [
    "K-disentQA",
    "SQA",
    "Instruct",
    "ASR",
    "Translation",
    "LSQA",
]

ROOT = Path(__file__).parent
DATA_PATH = ROOT / "data" / "leaderboard-data.json"
IMAGES_DIR = ROOT / "images"


def to_num(value: Any) -> float | None:
    try:
        num = float(value)
    except (TypeError, ValueError):
        return None
    return num


def average(values: list[float | None]) -> float | None:
    valid = [value for value in values if value is not None]
    if not valid:
        return None
    return sum(valid) / len(valid)


def compare_scores(left: float | None, right: float | None, lower_better: bool) -> int:
    if left is None and right is None:
        return 0
    if left is None:
        return 1
    if right is None:
        return -1
    if left == right:
        return 0
    if lower_better:
        return -1 if left < right else 1
    return -1 if left > right else 1


def ordered_tasks(tasks: list[dict[str, Any]]) -> list[dict[str, Any]]:
    def sort_key(task: dict[str, Any]) -> tuple[int, str]:
        try:
            task_index = TASK_ORDER.index(task["id"])
        except ValueError:
            task_index = 10**6
        return task_index, task["label"]

    return sorted(tasks, key=sort_key)


def dataset_ids(task: dict[str, Any]) -> list[str]:
    return [dataset["id"] for dataset in task.get("datasets", [])]


def metric_value(entry: dict[str, Any], task_id: str, dataset_id: str) -> float | None:
    dataset = entry.get("tasks", {}).get(task_id, {}).get(dataset_id)
    if not dataset:
        return None
    return to_num(dataset.get("value"))


def metric_display(entry: dict[str, Any], task_id: str, dataset_id: str) -> str:
    dataset = entry.get("tasks", {}).get(task_id, {}).get(dataset_id)
    if not dataset:
        return "-"
    if dataset.get("display") is not None:
        return str(dataset["display"])
    if dataset.get("value") is None:
        return "-"
    return str(dataset["value"])


def compute_task_overall(entry: dict[str, Any], task: dict[str, Any]) -> float | None:
    return average([metric_value(entry, task["id"], dataset_id) for dataset_id in dataset_ids(task)])


def normalize_task_scores(entries: list[dict[str, Any]], tasks: list[dict[str, Any]]) -> dict[str, dict[str, float] | None]:
    ranges: dict[str, dict[str, float] | None] = {}
    for task in tasks:
        values = [entry["task_overall"][task["id"]] for entry in entries if entry["task_overall"][task["id"]] is not None]
        if not values:
            ranges[task["id"]] = None
            continue
        ranges[task["id"]] = {"min": min(values), "max": max(values)}
    return ranges


def normalized_score(value: float | None, score_range: dict[str, float] | None, lower_better: bool) -> float | None:
    if value is None or score_range is None:
        return None
    if score_range["min"] == score_range["max"]:
        return 100.0
    if lower_better:
        return ((score_range["max"] - value) / (score_range["max"] - score_range["min"])) * 100
    return ((value - score_range["min"]) / (score_range["max"] - score_range["min"])) * 100


def enrich_entries(entries: list[dict[str, Any]], tasks: list[dict[str, Any]]) -> list[dict[str, Any]]:
    entries_with_task_overall = []
    for entry in entries:
        task_overall = {}
        for task in tasks:
            task_overall[task["id"]] = compute_task_overall(entry, task)
        entries_with_task_overall.append({**entry, "task_overall": task_overall})

    task_ranges = normalize_task_scores(entries_with_task_overall, tasks)
    enriched_entries = []
    for entry in entries_with_task_overall:
        normalized_task_scores = {}
        for task in tasks:
            normalized_task_scores[task["id"]] = normalized_score(
                entry["task_overall"][task["id"]],
                task_ranges[task["id"]],
                task["lowerBetter"],
            )
        enriched_entries.append(
            {
                **entry,
                "normalized_task_scores": normalized_task_scores,
                "overall": average([normalized_task_scores[task["id"]] for task in tasks]),
            }
        )
    return enriched_entries


def sort_overall(entries: list[dict[str, Any]]) -> list[dict[str, Any]]:
    sorted_entries = sorted(
        entries,
        key=cmp_to_key(lambda left, right: compare_scores(left["overall"], right["overall"], False)),
    )
    return [{**entry, "rank": index} for index, entry in enumerate(sorted_entries, start=1)]


def sort_task(entries: list[dict[str, Any]], task: dict[str, Any], dataset_id: str) -> list[dict[str, Any]]:
    def compare(left: dict[str, Any], right: dict[str, Any]) -> int:
        left_value = left["task_overall"][task["id"]] if dataset_id == "Overall" else metric_value(left, task["id"], dataset_id)
        right_value = right["task_overall"][task["id"]] if dataset_id == "Overall" else metric_value(right, task["id"], dataset_id)
        return compare_scores(left_value, right_value, task["lowerBetter"])

    sorted_entries = sorted(entries, key=cmp_to_key(compare))
    return [{**entry, "rank": index} for index, entry in enumerate(sorted_entries, start=1)]


def metric_class(lower_better: bool, value: float | None) -> str:
    if value is None:
        return "muted"
    return "metric-bad" if lower_better else "metric-good"


def fmt_score(value: float | None) -> str:
    return "-" if value is None else f"{value:.2f}"


def image_data_uri(path: Path) -> str:
    encoded = base64.b64encode(path.read_bytes()).decode("ascii")
    suffix = path.suffix.lower().lstrip(".") or "png"
    mime = "image/png" if suffix == "png" else f"image/{suffix}"
    return f"data:{mime};base64,{encoded}"


def load_payload() -> tuple[list[dict[str, Any]], list[dict[str, Any]]]:
    payload = json.loads(DATA_PATH.read_text(encoding="utf-8"))
    tasks = ordered_tasks(payload.get("tasks", []))
    entries = enrich_entries(payload.get("entries", []), tasks)
    return tasks, entries


TASKS, ENTRIES = load_payload()
TASK_MAP = {task["id"]: task for task in TASKS}
RANKED_OVERALL = sort_overall(ENTRIES)
BADGE_IMAGES = {
    1: image_data_uri(IMAGES_DIR / "1st.png"),
    2: image_data_uri(IMAGES_DIR / "2nd.png"),
    3: image_data_uri(IMAGES_DIR / "3rd.png"),
}
EXTERNAL_LINK_IMAGE = image_data_uri(IMAGES_DIR / "external-link.png")


def menu_choices() -> list[tuple[str, str]]:
    choices = [("Home\nOverall ranking", HOME_VIEW)]
    for task in TASKS:
        choices.append((f"{task['label']}\n{len(task['datasets'])} datasets", task["id"]))
    return choices


def dataset_choices(task_id: str) -> list[str]:
    task = TASK_MAP[task_id]
    return ["Overall", *[dataset["id"] for dataset in task.get("datasets", [])]]


def render_rank_strip(entries: list[dict[str, Any]]) -> str:
    cards = []
    for entry in entries[:12]:
        if entry["rank"] in BADGE_IMAGES:
            badge = f'<img class="rank-badge-image" src="{BADGE_IMAGES[entry["rank"]]}" alt="{entry["rank"]} place" />'
        else:
            badge = f'<span class="rank-badge">#{entry["rank"]}</span>'
        cards.append(
            f"""
            <article class="rank-pill">
              <div class="rank-badge-wrap">{badge}</div>
              <span class="rank-name">{escape(entry["rank_name"])}</span>
            </article>
            """
        )
    return f"""
    <section class="section-card card">
      <div class="section-head">
        <h3>Top Ranking</h3>
      </div>
      <div class="rank-strip-list">
        {"".join(cards)}
      </div>
    </section>
    """


def render_home_table(entries: list[dict[str, Any]]) -> str:
    header_top = [
        '<th rowspan="2" class="rank-col col-rank">Rank</th>',
        '<th rowspan="2" class="col-rankname">RankName</th>',
        '<th rowspan="2" class="col-model">Model</th>',
        '<th rowspan="2">URL</th>',
        '<th rowspan="2">Overall</th>',
        *[f'<th class="grouped" colspan="1">{escape(task["label"])}</th>' for task in TASKS],
    ]
    header_bottom = [f'<th>{escape(task["shortMetric"])}</th>' for task in TASKS]

    rows = []
    for entry in entries:
        url = entry.get("url") or ""
        if url:
            url_cell = (
                f'<a class="url-link" href="{escape(url)}" target="_blank" rel="noopener noreferrer" '
                f'aria-label="External link"><img src="{EXTERNAL_LINK_IMAGE}" width="5px" height="5px" alt="" /></a>'
            )
        else:
            url_cell = "-"

        task_cells = []
        for task in TASKS:
            value = entry["task_overall"][task["id"]]
            task_cells.append(f'<td><span class="{metric_class(task["lowerBetter"], value)}">{fmt_score(value)}</span></td>')

        rows.append(
            "<tr>"
            f'<td class="rank-col col-rank">{entry["rank"]}</td>'
            f'<td class="col-rankname">{escape(entry["rank_name"])}</td>'
            f'<td class="col-model">{escape(entry.get("model") or entry["rank_name"])}</td>'
            f"<td>{url_cell}</td>"
            f"<td>{fmt_score(entry['overall'])}</td>"
            f"{''.join(task_cells)}"
            "</tr>"
        )

    return f"""
    <section class="section-card card">
      <div class="section-head">
        <h3>Overall Leaderboard</h3>
      </div>
      <div class="table-scroll">
        <table>
          <colgroup>
            <col class="col-rank" />
            <col class="col-rankname" />
            <col class="col-model" />
            <col />
            <col />
            {"".join("<col />" for _ in TASKS)}
          </colgroup>
          <thead>
            <tr>{"".join(header_top)}</tr>
            <tr>{"".join(header_bottom)}</tr>
          </thead>
          <tbody>{"".join(rows)}</tbody>
        </table>
      </div>
    </section>
    """


def render_home() -> str:
    return f"{render_rank_strip(RANKED_OVERALL)}{render_home_table(RANKED_OVERALL)}"


def render_task_title(task: dict[str, Any]) -> str:
    return f"""
    <div class="section-head">
      <div>
        <h3 class="task-title">Task : {escape(task["label"])}</h3>
      </div>
    </div>
    """


def render_task_table(task: dict[str, Any], dataset_id: str) -> str:
    ranked_entries = sort_task(ENTRIES, task, dataset_id)
    active_label = "Overall"
    if dataset_id != "Overall":
        active_label = next(
            (dataset["label"] for dataset in task["datasets"] if dataset["id"] == dataset_id),
            dataset_id,
        )

    rows = []
    for entry in ranked_entries:
        numeric_value = entry["task_overall"][task["id"]] if dataset_id == "Overall" else metric_value(entry, task["id"], dataset_id)
        display_value = fmt_score(numeric_value) if dataset_id == "Overall" else metric_display(entry, task["id"], dataset_id)
        rows.append(
            "<tr>"
            f'<td class="rank-col col-rank">{entry["rank"]}</td>'
            f'<td class="col-rankname">{escape(entry["rank_name"])}</td>'
            f'<td class="col-model">{escape(entry.get("model") or entry["rank_name"])}</td>'
            f'<td><span class="{metric_class(task["lowerBetter"], numeric_value)}">{escape(display_value)}</span></td>'
            "</tr>"
        )

    return f"""
    <div id="taskTableMount">
      <div class="table-scroll">
        <table class="task-performance-table">
          <colgroup>
            <col class="col-rank" />
            <col class="col-rankname" />
            <col class="col-model" />
            <col />
          </colgroup>
          <thead>
            <tr>
              <th class="rank-col col-rank">Rank</th>
              <th class="col-rankname">RankName</th>
              <th class="col-model">Model</th>
              <th>{escape(active_label)}</th>
            </tr>
          </thead>
          <tbody>{"".join(rows)}</tbody>
        </table>
      </div>
    </div>
    """


def update_view(active_view: str, current_dataset: str | None) -> tuple[Any, Any, str, str, str]:
    if active_view == HOME_VIEW:
        first_task = TASKS[0]
        return (
            gr.update(visible=True),
            gr.update(visible="hidden"),
            gr.update(choices=dataset_choices(first_task["id"]), value="Overall"),
            render_task_title(first_task),
            first_task["metricLabel"],
            render_task_table(first_task, "Overall"),
        )

    task = TASK_MAP[active_view]
    choices = dataset_choices(active_view)
    dataset_id = current_dataset if current_dataset in choices else "Overall"
    return (
        gr.update(visible="hidden"),
        gr.update(visible=True),
        gr.update(choices=choices, value=dataset_id),
        render_task_title(task),
        task["metricLabel"],
        render_task_table(task, dataset_id),
    )


CUSTOM_CSS = """
@import url('https://fonts.googleapis.com/css2?family=Space+Grotesk:wght@400;500;700&family=Noto+Sans+KR:wght@400;500;700&display=swap');

:root {
  --bg: #f2f4f8;
  --bg-strong: #dde5f2;
  --panel: rgba(255, 255, 255, 0.84);
  --panel-strong: #ffffff;
  --text: #0c1730;
  --muted: #66748b;
  --line: rgba(12, 23, 48, 0.14);
  --primary: #0e3a8a;
  --accent: #c56b12;
  --pending: #74674a;
  --pending-bg: #f4ecd8;
  --success: #0c8f61;
  --danger: #b8612f;
  --shadow: 0 14px 38px rgba(12, 23, 48, 0.08);
}

html, body, .gradio-container {
  margin: 0 !important;
  min-height: 100vh;
  font-family: "Noto Sans KR", "Space Grotesk", sans-serif !important;
  color: var(--text);
  background:
    radial-gradient(circle at 15% 0%, #dce8ff 0%, transparent 32%),
    radial-gradient(circle at 95% 5%, #ffe4cf 0%, transparent 24%),
    linear-gradient(180deg, #f8fbff 0%, #eef2f7 100%) !important;
}

.gradio-container {
  max-width: 100% !important;
}

.app-root {
  position: relative;
  max-width: 1440px;
  margin: 0 auto;
  padding: 24px 20px 56px;
  overflow-x: auto;
  -webkit-overflow-scrolling: touch;
}

.bg-orb {
  position: fixed;
  border-radius: 999px;
  filter: blur(70px);
  opacity: 0.45;
  pointer-events: none;
  z-index: 0;
}

.orb-1 {
  width: 280px;
  height: 280px;
  background: #a7c5ff;
  top: -110px;
  left: -80px;
}

.orb-2 {
  width: 240px;
  height: 240px;
  background: #ffd2a4;
  top: -70px;
  right: -60px;
}

.layout-row {
  position: relative;
  z-index: 1;
  flex-wrap: nowrap !important;
  gap: 24px;
  align-items: flex-start;
  min-width: 1180px;
}

.sidebar-panel {
  flex: 0 0 260px !important;
  width: 260px !important;
  min-width: 260px !important;
  max-width: 260px !important;
  position: sticky;
  top: 20px;
  align-self: start;
  border: 1px solid var(--line);
  border-radius: 26px;
  background: rgba(7, 17, 40, 0.95);
  color: #ffffff;
  padding: 22px 18px;
  box-shadow: var(--shadow);
}

.sidebar-head,
.sidebar-head * {
  color: #ffffff !important;
}

.sidebar-head {
  margin-bottom: 26px;
}

.sidebar-kicker, .kicker {
  margin: 0;
  letter-spacing: 0.12em;
  text-transform: uppercase;
  font-size: 12px;
}

.sidebar-kicker {
  color: rgba(255, 255, 255, 0.82) !important;
}


.sidebar-head h1 {
  color: #ffffff !important;
}

.sidebar-head h1 {
  margin: 8px 0 0;
  font-size: 34px;
  line-height: 0.95;
  color: #ffffff;
}

.content-panel {
  flex: 1 1 auto !important;
  min-width: 0;
}

.hero {
  margin-bottom: 20px;
}

.hero .kicker {
  font-size: 14px !important;
  letter-spacing: 0.18em;
  color: var(--muted) !important;
}

.hero-topline {
  display: flex;
  align-items: flex-start;
  justify-content: space-between;
  gap: 20px;
  text-align: center;
}

.hero-topline > div {
  flex: 1;
}

.hero h2 {
  margin: 8px 0 6px;
  font-family: "Space Grotesk", sans-serif;
  font-size: clamp(56px, 7vw, 80px) !important;
  line-height: 1.02;
  color: var(--text);
  font-weight: 700;
}

.desc {
  margin: 0 auto;
  max-width: 760px;
  color: var(--muted);
  font-size: 17px;
}

.card {
  background: var(--panel);
  border: 1px solid var(--line);
  border-radius: 24px;
  box-shadow: var(--shadow);
  backdrop-filter: blur(8px);
}

.section-card {
  padding: 18px;
}

.section-head {
  display: flex;
  align-items: baseline;
  justify-content: space-between;
  gap: 12px;
  margin-bottom: 12px;
}

.section-head h3,
.task-title {
  margin: 0;
  font-size: 28px;
  font-family: "Space Grotesk", sans-serif;
}

.rank-strip-list {
  display: grid;
  grid-template-columns: repeat(6, minmax(120px, 1fr));
  gap: 10px;
}

.rank-pill {
  border: 1px solid var(--line);
  border-radius: 16px;
  padding: 12px;
  background: linear-gradient(135deg, #ffffff 0%, var(--bg-strong) 100%);
}

.rank-badge-wrap {
  min-height: 28px;
}

.rank-badge {
  display: inline-flex !important;
  align-items: center;
  justify-content: center;
  min-width: 34px;
  height: 34px;
  padding: 0 10px;
  border-radius: 999px;
  background: var(--primary);
  color: #ffffff !important;
  font-size: 13px;
  font-weight: 700;
  line-height: 1;
}

.rank-badge-image {
  display: block;
  width: auto;
  height: 28px;
}

.rank-name {
  display: block;
  margin-top: 8px;
  font-weight: 700;
  color: var(--text) !important;
}

.table-scroll {
  overflow-x: auto;
  overflow-y: visible;
  border-radius: 0;
  border: 1px solid #ffffff !important;
  background: var(--panel-strong);
}

table {
  width: 100%;
  min-width: 1080px;
  border-collapse: collapse;
  border: 1px solid #ffffff !important;
}

.task-performance-table {
  table-layout: fixed;
}

thead th {
  background: #e8edf6;
  border-bottom: 1px solid #ffffff !important;
  white-space: nowrap;
}

thead tr:first-child th.grouped {
  text-align: center;
}

th, td {
  padding: 12px 14px;
  text-align: left;
  border-bottom: 1px solid #ffffff !important;
  border-right: 1px solid #ffffff !important;
  font-size: 14px;
}

th:first-child,
td:first-child {
  border-left: 1px solid #ffffff !important;
}

th:last-child,
td:last-child {
  border-right: 0;
}

.table-scroll table,
.table-scroll thead,
.table-scroll tbody,
.table-scroll tr,
.table-scroll th,
.table-scroll td {
  border-color: #ffffff !important;
}

tbody tr:hover {
  background: #f8fbff;
}

.rank-col {
  font-family: "Space Grotesk", sans-serif;
  font-weight: 700;
  width: 72px;
}

.col-rank {
  width: 96px;
}

.col-rankname {
  width: 240px;
}

.col-model {
  width: 320px;
}

.url-link {
  display: inline-flex;
  align-items: center;
  justify-content: center;
  width: 24px;
  height: 24px;
}

.url-link img {
  display: block;
  width: 18px;
  height: 18px;
}

.metric-good {
  color: var(--success);
  font-weight: 700;
}

.metric-bad {
  color: var(--danger);
  font-weight: 700;
}

.muted {
  color: var(--muted);
}

.task-filters {
  display: flex;
  flex-wrap: nowrap;
  align-items: flex-start;
  gap: 12px;
  margin-top: 12px;
  width: 100%;
  overflow-x: auto;
  -webkit-overflow-scrolling: touch;
  padding-bottom: 6px;
  background: transparent !important;
  border: 0 !important;
  box-shadow: none !important;
}

.task-filters .dataset-wrap { flex: 0 0 620px; min-width: 620px; }
.task-filters .metric-wrap { flex: 0 0 340px; min-width: 340px; }

.task-filters .dataset-wrap,
.task-filters .metric-wrap,
.task-filters .dataset-wrap > div,
.task-filters .metric-wrap > div {
  background: transparent !important;
  border: 0 !important;
  box-shadow: none !important;
}

/* Keep Dataset / Metric columns pinned to the same top baseline. */
.task-filters .dataset-wrap,
.task-filters .metric-wrap {
  align-self: flex-start !important;
  justify-self: flex-start !important;
  margin-top: 0 !important;
  padding-top: 0 !important;
}

.task-view-shell .filter-title {
  margin: 0 0 6px 0 !important;
  font-size: 13px !important;
  color: var(--muted) !important;
  line-height: 1.2 !important;
}

#taskTableMount {
  margin-top: 18px;
}

.task-menu-radio {
  gap: 8px;
  background: transparent !important;
  border: 0 !important;
  box-shadow: none !important;
  padding: 0 !important;
}

.task-menu-radio > div,
.task-menu-radio .block,
.task-menu-radio .gradio-radio,
.task-menu-radio .form {
  background: transparent !important;
  border: 0 !important;
  box-shadow: none !important;
  padding: 0 !important;
}

.task-menu-radio label > span,
.task-menu-radio label > div,
.task-menu-radio label .wrap {
  white-space: pre-line !important;
}

.task-menu-radio label:has(input[type="radio"]) {
  width: 100%;
  margin: 0 !important;
  border: 1px solid rgba(255, 255, 255, 0.1) !important;
  background: rgba(255, 255, 255, 0.05) !important;
  color: #f5f7fb !important;
  border-radius: 14px !important;
  padding: 12px 14px !important;
  cursor: pointer !important;
  min-height: 72px;
  align-content: center;
  box-shadow: none !important;
  transition: background 0.18s ease, color 0.18s ease, border-color 0.18s ease;
}

.task-menu-radio label:has(input[type="radio"]):hover {
  background: rgba(255, 255, 255, 0.1) !important;
}

.task-menu-radio label:has(input[type="radio"]:checked) {
  background: linear-gradient(135deg, #fcf7eb 0%, #dfeaff 100%) !important;
  color: var(--text) !important;
  border-color: transparent !important;
}

.task-menu-radio input[type="radio"] {
  display: none !important;
}

.task-menu-radio label span,
.task-menu-radio label div,
.task-menu-radio label p {
  color: inherit !important;
}

.task-menu-radio label:has(input[type="radio"]) span,
.task-menu-radio label:has(input[type="radio"]) div {
  color: #f5f7fb !important;
}

.task-menu-radio label:has(input[type="radio"]:checked) span,
.task-menu-radio label:has(input[type="radio"]:checked) div {
  color: var(--text) !important;
}

.task-menu-radio .wrap,
.task-menu-radio label span:last-child {
  opacity: 0.72;
  font-size: 12px;
}

.task-view-shell .gradio-radio,
.task-view-shell .gradio-textbox {
  margin: 0 !important;
  min-width: 0 !important;
  background: transparent !important;
  border: 0 !important;
  box-shadow: none !important;
}

.task-view-shell .gradio-radio label,
.task-view-shell .gradio-textbox label {
  font-size: 13px !important;
  color: var(--muted) !important;
}

.task-view-shell .dataset-radio {
  background: transparent !important;
  border: 0 !important;
  box-shadow: none !important;
  padding: 0 !important;
  margin-left: 0 !important;
  margin-bottom: 12px !important;
  max-width: 100% !important;
  width: 100% !important;
}

.task-view-shell .dataset-radio > div,
.task-view-shell .dataset-radio .block,
.task-view-shell .dataset-radio .form {
  background: transparent !important;
  border: 0 !important;
  box-shadow: none !important;
  padding: 0 !important;
  display: grid !important;
  grid-template-columns: repeat(2, minmax(180px, 1fr)) !important;
  gap: 8px 10px !important;
  align-items: start !important;
  justify-content: start !important;
  min-height: 86px !important;
  align-content: start !important;
  max-width: 100% !important;
  width: 100% !important;
}

.task-view-shell .dataset-radio label:has(input[type="radio"]) {
  margin: 0 !important;
  border: 1px solid var(--line) !important;
  background: rgba(255, 255, 255, 0.88) !important;
  color: var(--text) !important;
  border-radius: 10px !important;
  padding: 0 12px !important;
  height: 40px !important;
  min-height: 40px !important;
  width: auto !important;
  display: flex !important;
  align-items: center !important;
  box-shadow: none !important;
  justify-content: flex-start !important;
  font-size: 14px !important;
  line-height: 1.2 !important;
}

.task-view-shell .dataset-radio label:has(input[type="radio"]:checked) {
  background: linear-gradient(135deg, #fcf7eb 0%, #dfeaff 100%) !important;
  border-color: rgba(14, 58, 138, 0.22) !important;
  font-weight: 700 !important;
}

.task-view-shell .dataset-radio input[type="radio"] {
  display: none !important;
}

.task-view-shell .metric-field,
.task-view-shell .dataset-field {
  align-self: flex-start !important;
  background: transparent !important;
}

/* Remove any residual top spacing on metric box so it aligns with Dataset. */
.task-view-shell .metric-wrap .metric-field,
.task-view-shell .metric-wrap .gradio-textbox,
.task-view-shell .metric-wrap .gradio-textbox > div,
.task-view-shell .metric-wrap .gradio-textbox .block,
.task-view-shell .metric-wrap .gradio-textbox .form,
.task-view-shell .metric-wrap .gradio-textbox .wrap {
  margin-top: 0 !important;
  padding-top: 0 !important;
}

.task-view-shell .metric-field > div,
.task-view-shell .dataset-field > div {
  background: transparent !important;
}

.task-view-shell .metric-field .gradio-textbox,
.task-view-shell .metric-field .gradio-textbox > div,
.task-view-shell .metric-field .gradio-textbox .block,
.task-view-shell .metric-field .gradio-textbox .form {
  background: transparent !important;
  border: 0 !important;
  box-shadow: none !important;
}

.task-view-shell .metric-field,
.task-view-shell .metric-field .gradio-textbox,
.task-view-shell .metric-field .gradio-textbox .wrap {
  width: 100% !important;
  max-width: none !important;
}

.task-view-shell .metric-field .gradio-textbox .wrap,
.task-view-shell .metric-field .gradio-textbox textarea,
.task-view-shell .metric-field .gradio-textbox input {
  height: 40px !important;
  min-height: 40px !important;
  width: 100% !important;
}

.task-view-shell input,
.task-view-shell textarea,
.task-view-shell .wrap-inner,
.task-view-shell button.secondary-down-arrow,
.task-view-shell .gradio-textbox .wrap {
  border-radius: 12px !important;
}

.task-view-shell .gradio-textbox .wrap,
.task-view-shell .gradio-textbox textarea,
.task-view-shell .gradio-textbox input {
  border: 1px solid var(--line) !important;
  background: rgba(255, 255, 255, 0.88) !important;
  color: var(--text) !important;
}

@media (max-width: 1280px) {
  .layout-row {
    min-width: 1120px;
  }

  .task-filters .dataset-wrap { flex-basis: 560px; min-width: 560px; }
  .task-filters .metric-wrap { flex-basis: 320px; min-width: 320px; }

  .task-view-shell .dataset-radio > div,
  .task-view-shell .dataset-radio .block,
  .task-view-shell .dataset-radio .form {
    grid-template-columns: repeat(2, minmax(150px, 1fr)) !important;
  }
}

@media (max-width: 980px) {
  .layout-row {
    min-width: 1040px;
  }

  .rank-strip-list {
    grid-template-columns: repeat(2, minmax(0, 1fr));
  }

  .task-view-shell .dataset-radio > div,
  .task-view-shell .dataset-radio .block,
  .task-view-shell .dataset-radio .form {
    grid-template-columns: repeat(2, minmax(140px, 1fr)) !important;
  }
}

@media (max-width: 720px) {
  .app-root {
    padding: 16px 14px 40px;
  }

  .hero-topline {
    flex-direction: column;
    align-items: center;
  }

  .hero h2 {
    font-size: 40px;
  }

  .rank-strip-list {
    grid-template-columns: 1fr;
  }

  .task-view-shell .dataset-radio > div,
  .task-view-shell .dataset-radio .block,
  .task-view-shell .dataset-radio .form {
    grid-template-columns: 1fr !important;
  }
}
"""


def build_app() -> gr.Blocks:
    with gr.Blocks(title="Ko-Speech-Eval Leaderboard", fill_width=True) as demo:
        with gr.Column(elem_classes=["app-root"]):
            gr.HTML('<div class="bg-orb orb-1"></div><div class="bg-orb orb-2"></div>')
            with gr.Row(elem_classes=["layout-row"]):
                with gr.Column(scale=0, min_width=260, elem_classes=["sidebar-panel"]):
                    gr.HTML(
                        """
                        <div class="sidebar-head">
                          <p class="sidebar-kicker">KoALa-bench</p>
                          <h1>Leaderboard</h1>
                          <a href="https://huggingface.co/datasets/scailaboratory/KoALA" target="_blank">
                                <img src="https://img.shields.io/badge/huggingface-FFD21E?style=flat&logo=huggingface&logoColor=white"/>
                          </a>
                          <a href="https://github.com/scai-research/KoALa-Bench?tab=readme-ov-file" target="_blank">
                                <img src="https://img.shields.io/badge/GiHub-181717?style=flat&logo=github&logoColor=white"/>
                          </a>
                        </div>
                        """
                    )
                    menu = gr.Radio(
                        choices=menu_choices(),
                        value=HOME_VIEW,
                        show_label=False,
                        container=False,
                        elem_classes=["task-menu", "task-menu-radio"],
                    )

                with gr.Column(scale=1, elem_classes=["content-panel"]):
                    gr.HTML(
                        """
                        <header class="hero">
                          <div class="hero-topline">
                            <div>
                              <p class="kicker">Korean Audio Language benchmark</p>
                              <h2>Leaderboard for KoALa</h2>
                            </div>
                          </div>
                        </header>
                        """
                    )

                    home_view = gr.HTML(render_home(), visible=True)

                    with gr.Column(visible="hidden", elem_classes=["task-view-shell", "section-card", "card"]) as task_view:
                        task_title = gr.HTML()
                        with gr.Row(elem_classes=["task-filters"]):
                            with gr.Column(scale=3, min_width=420, elem_classes=["dataset-wrap"]):
                                gr.HTML('<p class="filter-title">Dataset</p>')
                                dataset_dropdown = gr.Radio(
                                    choices=dataset_choices(TASKS[0]["id"]),
                                    value="Overall",
                                    show_label=False,
                                    elem_classes=["dataset-field", "dataset-radio"],
                                )
                            with gr.Column(scale=2, min_width=280, elem_classes=["metric-wrap"]):
                                gr.HTML('<p class="filter-title">Metric</p>')
                                metric_text = gr.Textbox(
                                    show_label=False,
                                    interactive=False,
                                    elem_classes=["metric-field"],
                                )
                        task_table = gr.HTML()

        menu.change(
            fn=update_view,
            inputs=[menu, dataset_dropdown],
            outputs=[home_view, task_view, dataset_dropdown, task_title, metric_text, task_table],
        )
        dataset_dropdown.change(
            fn=update_view,
            inputs=[menu, dataset_dropdown],
            outputs=[home_view, task_view, dataset_dropdown, task_title, metric_text, task_table],
        )
        demo.load(
            fn=update_view,
            inputs=[menu, dataset_dropdown],
            outputs=[home_view, task_view, dataset_dropdown, task_title, metric_text, task_table],
        )
    return demo


if __name__ == "__main__":
    app = build_app()
    app.launch(css=CUSTOM_CSS)