views.py

"""Gradio component builders for the leaderboard.

Each function creates one piece of the UI inside the surrounding Blocks/render
context. No data-path logic lives here (see ``data_loading.py``) and no copy
lives here (see ``website_texts.py``).
"""

from __future__ import annotations

import html
import re
from itertools import groupby
from pathlib import Path

import gradio as gr
import pandas as pd
from gradio_leaderboard import ColumnFilter, Leaderboard, SelectColumns

import website_texts
from constants import Constants, model_type_color, model_type_emoji
from data_loading import (
    BEYOND_SUBSET_LABELS,
    DATASET_SIZE_NOTE,
    BeyondSubset,
    LBContainer,
    Subset,
    load_leaderboard_csv,
    subset_name,
    unzip_png,
)

# --------------------------------------------------------------------------- #
# Full per-subset leaderboard table
# --------------------------------------------------------------------------- #

_IMPUTED_INFO = (
    "We impute the performance for models that cannot run on all datasets due to"
    " task or dataset size constraints. We impute with the performance of a"
    " default RandomForest. We add a postfix [X% IMPUTED] to the model if any"
    " results were imputed. The X% shows the percentage of datasets that were"
    " imputed. In general, imputation negatively represents the model"
    " performance, punishing the model for not being able to run on all datasets."
)


def make_leaderboard(lb: LBContainer) -> Leaderboard:
    df = lb.load_df()

    # -- Derived filter columns
    df["TypeFiler"] = df["TypeName"].apply(lambda m: f"{m} {model_type_emoji[m]}")
    df["Only Default"] = df["Model"].str.contains("(default)", regex=False)
    df["Only Tuned"] = df["Model"].str.contains("(tuned)", regex=False)
    df["Only Tuned + Ensembled"] = df["Model"].str.contains(
        "(tuned + ensembled)", regex=False
    ) | df["Model"].str.contains("AutoGluon", regex=False)

    filter_columns = [
        ColumnFilter("TypeFiler", type="checkboxgroup", label="🤖 Model Types"),
        ColumnFilter("Only Default", type="boolean", default=False),
        ColumnFilter("Only Tuned", type="boolean", default=False),
        ColumnFilter("Only Tuned + Ensembled", type="boolean", default=False),
    ]

    # -- Imputation column (only meaningful when some rows are imputed)
    if df["Imputed"].any():
        df["Imputed"] = df["Imputed"].replace({True: "Imputed", False: "Not Imputed"})
        filter_columns.append(
            ColumnFilter(
                "Imputed",
                type="checkboxgroup",
                label="(Not) Imputed Models",
                info=_IMPUTED_INFO,
            )
        )
    else:
        df = df.drop(columns=["Imputed (%) [⬇️]"])

    # -- Color model names by type so the table shares the overview's hue.
    def _color_model(row: pd.Series) -> str:
        color = model_type_color.get(row["TypeName"], "#dddddd")
        link = re.match(r"\[(.*?)\]\((.*?)\)(.*)", row["Model"])
        if link:
            text, url, rest = link.groups()
            return (
                f'<a href="{url}" target="_blank" rel="noopener" '
                f'style="color:{color};font-weight:600;">{text}</a>{rest}'
            )
        return f'<span style="color:{color};font-weight:600;">{row["Model"]}</span>'

    df["Model"] = df.apply(_color_model, axis=1)

    # -- Column datatypes (read straight from dtypes; transposing collapses to
    #    object and mistypes every numeric column as markdown).
    datatypes = []
    for dtype in df.dtypes:
        if pd.api.types.is_bool_dtype(dtype):  # check bool first (it is also "numeric")
            datatypes.append("bool")
        elif pd.api.types.is_numeric_dtype(dtype):
            datatypes.append("number")
        else:
            datatypes.append("markdown")

    return Leaderboard(
        elem_id=f"lb_{lb.subset.rel_path.replace('/', '_')}",
        value=df,
        datatype=datatypes,
        select_columns=SelectColumns(
            default_selection=list(df.columns),
            cant_deselect=["Type", "Model"],
            label="Select Columns to Display:",
        ),
        hide_columns=[
            "TypeName",
            "TypeFiler",
            "RefModel",
            "Only Default",
            "Only Tuned",
            "Only Tuned + Ensembled",
            "Imputed",
        ],
        search_columns=["Model", "TypeName"],
        filter_columns=filter_columns,
        bool_checkboxgroup_label="Custom Views (exclusive, only toggle one at a time):",
        height=800,
    )


# --------------------------------------------------------------------------- #
# Per-subset figures
# --------------------------------------------------------------------------- #


def make_overview_images(lb: LBContainer) -> None:
    name = subset_name(lb.subset)
    gr.Image(
        lb.image_path("tuning-impact-elo"),
        label=f"Leaderboard Overview [{name}]",
        show_label=True,
        height=500,
        show_share_button=True,
    )
    with gr.Row():
        with gr.Column(scale=1):
            gr.Image(
                value=lb.image_path("pareto_front_improvability_vs_time_infer"),
                label=f"Inference Time Pareto Front [{name}]",
                height=400,
                show_label=True,
                show_share_button=True,
            )
        with gr.Column(scale=1):
            gr.Image(
                value=lb.image_path("pareto_n_configs_imp"),
                label=f"Tuning Trajectories [{name}]",
                height=400,
                show_label=True,
                show_share_button=True,
            )


def make_winrate_image(lb: LBContainer) -> None:
    # Uses ``lb.name`` (the caller-supplied subset name) so it is benchmark-agnostic and reusable
    # across both the TabArena and BeyondArena tabs.
    gr.Image(
        lb.image_path("winrate_matrix"),
        label=f"Win-rate Matrix [{lb.name}]",
        show_label=True,
        height=800,
        show_share_button=True,
    )


# --------------------------------------------------------------------------- #
# Cross-subset overview (Elo heatmap, rendered as HTML for links + grouping)
# --------------------------------------------------------------------------- #

# Columns of the overview: (label, group, subset). Always imputation=yes/splits=all.
_OVERVIEW_COLUMNS: list[tuple[str, str, Subset]] = [
    ("Overall", "overall", Subset(tasks="all", datasets="all")),
    ("Class.", "task", Subset(tasks="classification", datasets="all")),
    ("Regr.", "task", Subset(tasks="regression", datasets="all")),
    ("Binary", "task", Subset(tasks="binary", datasets="all")),
    ("Multi.", "task", Subset(tasks="multiclass", datasets="all")),
    ("Small", "size", Subset(tasks="all", datasets="small")),
    ("Medium", "size", Subset(tasks="all", datasets="medium")),
]
_GROUP_TITLE = {"task": "By Task", "size": "By Dataset Size"}
# Hover tooltips for the overview's subset column headers (rendered as a native `title=`).
# Size definitions are reused from DATASET_SIZE_NOTE so they can't drift from the subset blurbs.
_COLUMN_TOOLTIPS: dict[str, str] = {
    "Overall": "All tasks across every dataset size — the headline ranking.",
    "Class.": "Classification tasks only (binary + multiclass).",
    "Regr.": "Regression tasks only.",
    "Binary": "Binary classification tasks only.",
    "Multi.": "Multiclass classification tasks only.",
    "Small": DATASET_SIZE_NOTE["small"],
    "Medium": DATASET_SIZE_NOTE["medium"],
}
# Top-3 per column get a medal in a fixed-width slot reserved in every number cell, so the
# numbers line up whether or not a medal is present (see `.ta-medal`/`.ta-val` in `_OVERVIEW_CSS`).
_MEDALS = {1: "🥇", 2: "🥈", 3: "🥉"}
_VARIANT_RE = re.compile(r"\((default|tuned \+ ensembled|tuned)\)")

# Metrics selectable in the overview: label -> (csv column, higher_is_better, formatter).
# Order = display order in the selector; the two headline metrics come first.
_OVERVIEW_METRIC_SPECS: dict[str, tuple] = {
    "Elo": ("Elo [⬆️]", True, lambda v: str(int(round(v)))),
    "Improvability (%)": ("Improvability (%) [⬇️]", False, lambda v: f"{v:.1f}"),
    "Score": ("Score [⬆️]", True, lambda v: f"{v:.3f}"),
    "Average Rank": ("Rank [⬇️]", False, lambda v: f"{v:.2f}"),
    "Harmonic Rank": ("Harmonic Rank [⬇️]", False, lambda v: f"{v:.2f}"),
}
OVERVIEW_METRIC_CHOICES = list(_OVERVIEW_METRIC_SPECS)

# One-line TL;DR shown next to the overview metric selector.
OVERVIEW_METRIC_TLDR = {
    "Elo": "Pairwise win-rate rating (a 400-point gap ≈ a 91% win rate). Higher is better.",
    "Improvability (%)": "How much lower the best model's error is than this one's, per dataset. Lower is better.",
    "Score": "Error rescaled per dataset to 1 (best) … 0 (median), then averaged. Higher is better.",
    "Average Rank": "The model's mean rank across datasets. Lower is better.",
    "Harmonic Rank": "Harmonic mean of per-dataset ranks — rewards being excellent on some datasets. Lower is better.",
}


def _parse_model(model: str) -> tuple[str, str, str | None]:
    """Split a Model cell into (base name, variant, url)."""
    link = re.match(r"\[(.*?)\]\((.*?)\)", model)
    text, url = (link.group(1), link.group(2)) if link else (model, None)
    text = text.split("[")[0].strip()  # drop any [X% IMPUTED] tag
    variant_match = _VARIANT_RE.search(text)
    variant = variant_match.group(1) if variant_match else ""
    base = _VARIANT_RE.sub("", text).strip()
    return base, variant, url


def _subset_best(df: pd.DataFrame, column: str, higher_is_better: bool) -> pd.DataFrame:
    """Best variant per model in a subset for `column`; excludes reference pipelines."""
    df = df[~df["TypeName"].isin([Constants.reference])].copy()
    df = df.dropna(subset=[column])
    parsed = df["Model"].map(_parse_model)
    df["base"] = [p[0] for p in parsed]
    df["variant"] = [p[1] for p in parsed]
    df["url"] = [p[2] for p in parsed]
    df["imputed"] = df["Imputed"].astype(bool) if "Imputed" in df.columns else False
    df["verified"] = df["Verified"] if "Verified" in df.columns else ""
    grouped = df.groupby("base")[column]
    best = df.loc[grouped.idxmax() if higher_is_better else grouped.idxmin()]
    return best[
        ["base", "TypeName", "Type", "variant", "url", column, "verified", "imputed"]
    ].rename(columns={column: "val"})


def _overview_th(label: str, *, rowspan: int | None = None) -> str:
    """A subset column header `<th>`; gets a hover tooltip + 'help' affordance when defined."""
    attrs = f' rowspan="{rowspan}"' if rowspan else ""
    tooltip = _COLUMN_TOOLTIPS.get(label)
    if tooltip:
        attrs += f' class="ta-th-info" title="{html.escape(tooltip, quote=True)}"'
    return f"<th{attrs}>{label}</th>"


def _interp_color(frac: float) -> str:
    """Map 0 (best) .. 1 (worst) to a green->olive->red hex (readable on dark bg)."""
    stops = [(0.0, (28, 120, 62)), (0.5, (138, 122, 36)), (1.0, (160, 58, 58))]
    frac = max(0.0, min(1.0, frac))
    for (f0, c0), (f1, c1) in zip(stops, stops[1:]):
        if frac <= f1:
            t = 0 if f1 == f0 else (frac - f0) / (f1 - f0)
            r, g, b = (round(a + (b_ - a) * t) for a, b_ in zip(c0, c1))
            return f"#{r:02x}{g:02x}{b:02x}"
    return "#a03a3a"


_OVERVIEW_CSS = """
<style>
.ta-overview { border-collapse: collapse; width: 100%; font-size: 1.08em; }
.ta-overview th, .ta-overview td { padding: 5px 9px; text-align: center; border: 1px solid #ffffff14; }
.ta-overview thead th { background: #1b1b22; font-weight: 600; position: sticky; z-index: 5; box-shadow: inset 0 1px 0 #ffffff14, inset 0 -1px 0 #ffffff14; }
/* Subset column headers carry an explanatory tooltip (title=); hint it with a help cursor + dotted underline. */
.ta-overview thead th.ta-th-info { cursor: help; text-decoration: underline; text-decoration-style: dotted; text-decoration-color: #ffffff66; text-underline-offset: 3px; }
.ta-overview thead tr:first-child th { top: 0; }
.ta-overview thead tr:nth-child(2) th { top: 33px; }
.ta-overview .ta-group-h { border-bottom: 2px solid #ffffff33; }
.ta-overview td.ta-model-cell { text-align: left; white-space: nowrap; }
.ta-overview .ta-variant { opacity: 0.5; font-weight: 400; font-size: 0.9em; }
/* Numbers match the rest of the table text in size; bold keeps them emphasized. */
.ta-overview td.ta-num { font-weight: 600; }
/* Reserve a fixed-width medal slot in every number cell so the digits line up in their
   own sub-column whether or not a top-3 medal is present. */
.ta-overview td.ta-num .ta-cell { display: inline-flex; align-items: baseline; }
.ta-overview td.ta-num .ta-medal { flex: 0 0 1.45em; width: 1.45em; text-align: left; font-size: 0.78em; }
.ta-overview td.ta-na { color: #777; }
.ta-overview td.ta-model-cell a { text-decoration: underline; text-decoration-style: dotted; text-underline-offset: 3px; }
.ta-overview td.ta-model-cell a:hover { text-decoration-style: solid; }
.ta-link-icon { font-size: 0.78em; opacity: 0.65; margin-left: 2px; }
.ta-imp { color: #e6c14d; font-weight: 700; margin-left: 1px; }
.ta-verified { font-size: 0.85em; }
.ta-pill { padding: 1px 7px; border-radius: 999px; font-size: 0.95em; }
.ta-scroll {
    overflow: auto;
    max-height: 680px;
    border: 1px solid #ffffff1f;
    border-radius: 10px;
    scrollbar-width: thin;
    scrollbar-color: #ffffff3a transparent;
    scrollbar-gutter: stable;
}
.ta-scroll::-webkit-scrollbar { width: 11px; height: 11px; }
.ta-scroll::-webkit-scrollbar-track { background: transparent; }
.ta-scroll::-webkit-scrollbar-thumb { background: #ffffff33; border-radius: 8px; border: 3px solid transparent; background-clip: content-box; }
.ta-scroll::-webkit-scrollbar-thumb:hover { background: #ffffff5c; background-clip: content-box; }
.ta-scroll::-webkit-scrollbar-corner { background: transparent; }
.ta-cap { font-size: 0.85em; opacity: 0.8; margin: 6px 0 4px 0; }
.ta-legend { margin: 0 0 10px 0; }
</style>
"""


def type_legend_html(include_reference: bool = True) -> str:
    """A small legend explaining the model-type symbols (shared across tables).

    Baseline and Other share one color, so they are shown as a single entry.
    `include_reference` is False for the overview, which excludes reference pipelines.
    """
    e = model_type_emoji
    entries = [
        (Constants.foundational, e[Constants.foundational], "Foundation Model"),
        (Constants.neural_network, e[Constants.neural_network], "Neural Network"),
        (Constants.tree, e[Constants.tree], "Tree-based"),
        (Constants.baseline, f"{e[Constants.baseline]} {e[Constants.other]}", "Baseline / Other"),
    ]
    if include_reference:
        entries.append((Constants.reference, e[Constants.reference], "Reference Pipeline"))
    chips = []
    for type_name, symbols, label in entries:
        color = model_type_color.get(type_name, "#9e9e9e")
        chips.append(
            f'<span style="display:inline-block;margin:2px 12px 2px 0;white-space:nowrap;">'
            f'<span class="ta-pill" style="background:{color}22;color:{color};border:1px solid {color}66;">{symbols}</span> '
            f'<span style="color:{color};font-size:0.85em;">{label}</span></span>'
        )
    return f'{_OVERVIEW_CSS}<div class="ta-legend">{"".join(chips)}</div>'


def make_cross_subset_overview(data_root: Path, metric: str = "Elo") -> gr.HTML:
    """Heatmap of the best `metric` per model (rows) and subset (columns)."""
    if metric not in _OVERVIEW_METRIC_SPECS:
        metric = "Elo"
    column, higher_is_better, fmt = _OVERVIEW_METRIC_SPECS[metric]

    val_by_col: dict[str, dict[str, float]] = {}
    imp_by_col: dict[str, dict[str, bool]] = {}
    meta: dict[str, dict] = {}
    present: list[tuple[str, str]] = []  # (label, group)

    for label, group, subset in _OVERVIEW_COLUMNS:
        path = Path(data_root) / subset.rel_path / "website_leaderboard.csv"
        if not path.exists():
            continue
        best = _subset_best(load_leaderboard_csv(str(path.resolve())), column, higher_is_better)
        val_by_col[label] = dict(zip(best["base"], best["val"]))
        imp_by_col[label] = dict(zip(best["base"], best["imputed"]))
        present.append((label, group))
        for _, row in best.iterrows():
            meta.setdefault(
                row["base"],
                {
                    "type_name": row["TypeName"],
                    "emoji": row["Type"],
                    "variant": row["variant"],
                    "url": row["url"],
                    "verified": row["verified"],
                },
            )

    if not present:
        return gr.HTML("<p>No overview data available.</p>")

    sort_label = present[0][0]
    worst_sort = float("-inf") if higher_is_better else float("inf")
    bases = sorted(
        meta, key=lambda b: val_by_col[sort_label].get(b, worst_sort), reverse=higher_is_better
    )

    rank_by_col, bounds = {}, {}
    for label, _ in present:
        col = val_by_col[label]
        ranked = sorted(col, key=lambda b: col[b], reverse=higher_is_better)
        rank_by_col[label] = {b: i + 1 for i, b in enumerate(ranked[:3])}
        bounds[label] = (min(col.values()), max(col.values())) if col else (0.0, 1.0)

    # -- Grouped header (two rows)
    row1 = ['<th rowspan="2">Type</th>', '<th rowspan="2">Model</th>']
    row2 = []
    for group, items in groupby(present, key=lambda x: x[1]):
        items = list(items)
        if group == "overall":
            for label, _ in items:
                row1.append(_overview_th(label, rowspan=2))
        else:
            row1.append(f'<th colspan="{len(items)}" class="ta-group-h">{_GROUP_TITLE[group]}</th>')
            row2.extend(_overview_th(label) for label, _ in items)
    header = f"<thead><tr>{''.join(row1)}</tr><tr>{''.join(row2)}</tr></thead>"

    # -- Body
    body = []
    for base in bases:
        m = meta[base]
        color = model_type_color.get(m["type_name"], "#9e9e9e")
        name = html.escape(base)
        if m["variant"]:
            name += f' <span class="ta-variant">({html.escape(m["variant"])})</span>'
        if m.get("verified") == "✔️":
            name += ' <span class="ta-verified" title="Verified implementation">✔️</span>'
        if m["url"]:
            name_html = (
                f'<a href="{html.escape(m["url"])}" target="_blank" rel="noopener" '
                f'style="color:{color};font-weight:600;">{name}<span class="ta-link-icon">↗</span></a>'
            )
        else:
            name_html = f'<span style="color:{color};font-weight:600;">{name}</span>'
        cells = [
            f'<td><span class="ta-pill" style="background:{color}22;color:{color};'
            f'border:1px solid {color}66;">{m["emoji"]}</span></td>',
            f'<td class="ta-model-cell">{name_html}</td>',
        ]
        for label, _ in present:
            val = val_by_col[label].get(base)
            if val is None:
                cells.append('<td class="ta-na">–</td>')
                continue
            lo, hi = bounds[label]
            if hi <= lo:
                frac_best = 0.5
            else:
                frac_best = (val - lo) / (hi - lo) if higher_is_better else (hi - val) / (hi - lo)
            bg = _interp_color(1 - frac_best)
            medal = _MEDALS.get(rank_by_col[label].get(base), "")
            imp = (
                '<sup class="ta-imp" title="Score is (partly) imputed">*</sup>'
                if imp_by_col[label].get(base)
                else ""
            )
            cells.append(
                f'<td class="ta-num" style="background:{bg};color:#f7f7f7;">'
                f'<span class="ta-cell"><span class="ta-medal">{medal}</span>'
                f'<span class="ta-val">{fmt(val)}{imp}</span></span></td>'
            )
        body.append(f"<tr>{''.join(cells)}</tr>")

    direction = "Higher is better" if higher_is_better else "Lower is better"
    caption = (
        f'<div class="ta-cap">Best <b>{html.escape(metric)}</b> per model across subsets '
        f"(with imputation, all repeats). {direction}; 🥇🥈🥉 mark the top 3 in each column. "
        "Each model shows its best-performing variant.</div>"
        '<div class="ta-cap ta-cap-legend">✔️ = verified implementation &nbsp;·&nbsp; '
        '<span class="ta-imp">*</span> = (partly) imputed score &nbsp;·&nbsp; '
        "💡 Click any <u>underlined</u> model name (↗) to open its paper or code.</div>"
    )
    table = f'<table class="ta-overview">{header}<tbody>{"".join(body)}</tbody></table>'
    return gr.HTML(
        f'{type_legend_html(include_reference=False)}{caption}<div class="ta-scroll">{table}</div>',
        elem_classes="ta-overview-block",
    )


# --------------------------------------------------------------------------- #
# Metric reference and agentic guide
# --------------------------------------------------------------------------- #


def make_metric_reference() -> None:
    """Compact column key — pills that reveal a single shared panel (one open at a time).

    Uses a CSS-only radio group (no JS): selecting a pill checks its hidden radio, which
    reveals just that definition in the shared panel below; the ✕ rechecks a hidden
    "none" radio to close.
    """
    radios = ['<input type="radio" name="ta-mk" id="ta-mk-none" class="ta-mk-radio" checked>']
    labels, defs, rules = [], [], []
    for i, metric in enumerate(website_texts.METRICS):
        rid = f"ta-mk-{i}"
        name = html.escape(metric["name"])
        definition = html.escape(f"{metric['details']}  ·  Why we use it: {metric['why']}")
        radios.append(f'<input type="radio" name="ta-mk" id="{rid}" class="ta-mk-radio">')
        labels.append(f'<label for="{rid}" class="ta-mk-chip">{name}</label>')
        defs.append(
            f'<div class="ta-mk-def" id="{rid}-def">'
            f'<label for="ta-mk-none" class="ta-mk-close" title="Close">✕</label>{definition}</div>'
        )
        rules.append(f"#{rid}:checked ~ .ta-mk-panel #{rid}-def{{display:block}}")
        rules.append(
            f"#{rid}:checked ~ .ta-mk-chips label[for='{rid}']"
            "{background:rgba(110,140,245,0.22);border-color:rgba(110,140,245,0.9);color:#cdd7ff}"
        )
    gr.HTML(
        "<style>" + "".join(rules) + "</style>"
        '<div class="ta-metric-key">' + "".join(radios)
        + '<div class="ta-mk-head">'
        '<span class="ta-mk-title">📊 Column key</span>'
        '<span class="ta-mk-hint">ranked by the <b>Elo</b> aggregation · click any column for its definition</span>'
        "</div>"
        + '<div class="ta-mk-chips">' + "".join(labels) + "</div>"
        + '<div class="ta-mk-panel">' + "".join(defs) + "</div>"
        + "</div>"
    )


def make_agentic_guide() -> None:
    gr.Markdown(website_texts.AGENTIC_GUIDE, elem_classes="markdown-text-box")


def make_hero_stats(data_root: Path) -> gr.HTML:
    """A compact strip of headline-fact cards shown above the info boxes."""
    lb = LBContainer(data_root, Subset(), "")
    n_datasets = lb.n_datasets or "—"
    df = lb.load_df()
    df = df[~df["TypeName"].isin([Constants.reference])]
    n_models = len({_parse_model(m)[0] for m in df["Model"]})
    paper = "https://tabarena.ai/paper-tabular-ml-iid-study"
    code = "https://tabarena.ai/code"

    cards = [
        (
            "🧾",
            f"{n_datasets} datasets",
            f'curated from 1,053 (<a href="{paper}" target="_blank" rel="noopener">see paper</a>)',
        ),
        ("🤖", f"{n_models}+ models", "state-of-the-art, each tuned to its peak"),
        (
            "✅",
            "Open-source",
            f'verified implementations (<a href="{code}" target="_blank" rel="noopener">see code</a>)',
        ),
        ("⚖️", "Scientifically rigorous", "strong validation, reproducible"),
    ]
    chips = "".join(
        f'<div class="ta-card"><div class="ta-card-ico">{ico}</div>'
        f'<div class="ta-card-body"><div class="ta-card-num">{num}</div>'
        f'<div class="ta-card-lbl">{lbl}</div></div></div>'
        for ico, num, lbl in cards
    )
    # Tagline rendered as a full-width card in the same group, above the stat boxes.
    intro = " ".join(website_texts.INTRODUCTION_TEXT.split())
    intro = re.sub(r"\*\*(.+?)\*\*", r"<b>\1</b>", intro)
    return gr.HTML(f'<div class="ta-hero"><div class="ta-intro">{intro}</div>{chips}</div>')


# --------------------------------------------------------------------------- #
# BeyondArena components
#
# BeyondArena reuses the per-subset leaderboard table (make_leaderboard) and the
# win-rate figure (make_winrate_image) unchanged; only the hero strip, the
# cross-subset overview (an image, not the TabArena HTML heatmap) and the
# per-subset figure set differ (no HPO tuning-trajectory figure — BeyondArena is
# evaluated on a single `core` protocol).
# --------------------------------------------------------------------------- #


# A distinct teal→violet accent for the BeyondArena hero, so its top reads clearly different from
# TabArena's neutral cards while staying in the same visual family (scoped to `.beyond-hero`).
_BEYOND_HERO_CSS = """
<style>
.beyond-hero .ta-card {
    border-color: #2dd4bf33;
    border-left: 3px solid #2dd4bf;
    background: linear-gradient(135deg, #2dd4bf1f, #7c5cf00d);
}
.beyond-hero .ta-card:hover { border-color: #2dd4bf99; }
.beyond-hero .ta-card-ico { color: #2dd4bf; }
.beyond-hero .ta-intro { border-left: 3px solid #7c5cf0; }
</style>
"""


def make_beyond_hero_stats(data_root: Path) -> gr.HTML:
    """A compact strip of headline-fact cards shown above the BeyondArena info boxes.

    Deliberately ordered / worded / colored differently from TabArena's hero: it leads with the
    beyond-IID identity and the curated subsets (the benchmark's focus) and uses a teal→violet accent.
    """
    lb = LBContainer(data_root, BeyondSubset("full"), "")
    n_datasets = lb.n_datasets or "—"
    df = lb.load_df()
    df = df[~df["TypeName"].isin([Constants.reference])]
    n_models = len({_parse_model(m)[0] for m in df["Model"]})
    # Curated subsets = every subset tab except the "full" (whole-benchmark) view.
    n_subsets = len(BEYOND_SUBSET_LABELS) - 1
    paper = "https://arxiv.org/abs/2606.30410"
    code = "https://tabarena.ai/code"

    cards = [
        ("🌍", "Beyond IID", "non-IID, temporal &amp; grouped tabular data"),
        ("🧩", f"{n_subsets} subsets", "curated subsets of the benchmark"),
        (
            "🧾",
            f"{n_datasets} datasets",
            f'across sizes &amp; dimensionalities (<a href="{paper}" target="_blank" rel="noopener">see paper</a>)',
        ),
        (
            "🤖",
            f"{n_models} models",
            f'tuned pipelines with preprocessing, beyond IID (<a href="{code}" target="_blank" rel="noopener">see code</a>)',
        ),
    ]
    chips = "".join(
        f'<div class="ta-card"><div class="ta-card-ico">{ico}</div>'
        f'<div class="ta-card-body"><div class="ta-card-num">{num}</div>'
        f'<div class="ta-card-lbl">{lbl}</div></div></div>'
        for ico, num, lbl in cards
    )
    intro = " ".join(website_texts.BEYOND_INTRODUCTION_TEXT.split())
    intro = re.sub(r"\*\*(.+?)\*\*", r"<b>\1</b>", intro)
    return gr.HTML(f'{_BEYOND_HERO_CSS}<div class="ta-hero beyond-hero"><div class="ta-intro">{intro}</div>{chips}</div>')


def make_beyond_overview_figure(data_root: Path) -> None:
    """The cross-subset overview: best Elo per model / per family across every subset.

    Unlike TabArena's HTML heatmap, BeyondArena's overview is the ``plot_subset_results`` image
    (``result_plots/per_model_elo`` + ``per_family_elo``). Missing images are skipped gracefully.
    """
    result_dir = Path(data_root) / "result_plots"
    images = [
        ("per_family_elo", "Best Elo per model type (family) across subsets"),
        ("per_model_elo", "Best Elo per model across subsets"),
    ]
    shown = False
    for name, label in images:
        if (result_dir / f"{name}.png").exists() or (result_dir / f"{name}.png.zip").exists():
            gr.Image(
                unzip_png(result_dir, name),
                label=label,
                show_label=True,
                height=520,
                show_share_button=True,
            )
            shown = True
    if not shown:
        gr.Markdown("_The cross-subset overview figure is not available yet._", elem_classes="markdown-text")


def make_beyond_subset_figures(lb: LBContainer) -> None:
    """Per-subset figures for BeyondArena: the Elo overview and the inference-time Pareto front."""
    name = lb.name
    gr.Image(
        lb.image_path("tuning-impact-elo"),
        label=f"Leaderboard Overview [{name}]",
        show_label=True,
        height=500,
        show_share_button=True,
    )
    gr.Image(
        value=lb.image_path("pareto_front_improvability_vs_time_infer"),
        label=f"Inference Time Pareto Front [{name}]",
        height=450,
        show_label=True,
        show_share_button=True,
    )