File size: 18,576 Bytes
4d6fb4f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
9a668d8
4d6fb4f
 
 
 
 
 
9a668d8
 
4d6fb4f
11a28fa
 
 
4d6fb4f
 
 
 
 
9a668d8
 
 
 
 
4d6fb4f
 
9a668d8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4d6fb4f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
9a668d8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4d6fb4f
9a668d8
 
4d6fb4f
 
 
 
 
 
 
 
 
9a668d8
 
 
 
 
 
4d6fb4f
9a668d8
4d6fb4f
 
 
 
 
 
 
 
9a668d8
4d6fb4f
 
 
 
 
 
 
9a668d8
 
 
 
 
 
 
4d6fb4f
 
 
 
9a668d8
 
 
 
4d6fb4f
 
218b520
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
11a28fa
 
 
 
 
 
 
 
 
 
 
4d6fb4f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
9a668d8
 
 
 
 
 
 
 
 
4d6fb4f
 
11a28fa
4d6fb4f
96e822b
 
 
 
 
 
 
 
 
 
 
 
4d6fb4f
 
 
 
 
 
9a668d8
 
4d6fb4f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
9a668d8
 
 
 
4d6fb4f
 
 
 
11a28fa
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
9a668d8
 
11a28fa
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
"""Data loading and subset configuration for the TabArena leaderboard.

This module owns everything about *where* leaderboard artifacts live and *how*
they are read. Layout (Gradio components) lives in ``views.py`` and ``pages.py``;
user-facing copy lives in ``website_texts.py``.

Performance note: the website optimizes for fast first paint. CSVs are tiny and
cached (:func:`load_leaderboard_csv`); the large per-subset PNGs are only
unzipped on demand (:meth:`LBContainer.image_path`) and only for the subset the
user is currently viewing.
"""

from __future__ import annotations

import re
import zipfile
from collections.abc import Iterable
from dataclasses import dataclass, field
from functools import lru_cache
from pathlib import Path

import pandas as pd

from constants import Constants

DATA_DIR = Path(__file__).parent / "data"
# BeyondArena artifacts live under their own root (see
# scripts/run_generate_beyondarena_website_artifacts.py in the tabarena repo).
BEYOND_DATA_DIR = Path(__file__).parent / "data_beyondarena"


# --------------------------------------------------------------------------- #
# Subset axes
#
# A leaderboard "subset" is one cell of a 5-axis grid. Three axes are *view
# modifiers* surfaced as controls (entrants, imputation, splits); two are
# *content subsets* surfaced as tab bars (tasks, datasets). Keeping the axis
# definitions here (as data, not as if/elif chains) means adding or reordering a
# subset is a one-line edit. The first value of each axis is its default.
# --------------------------------------------------------------------------- #

# Who competes. Every leaderboard number is relative to the field: Elo is pairwise over
# the participants and improvability is measured against the best of them, so each pool is
# its own evaluation with its own artifacts rather than a filter over a shared table.
#
# Models always compete. Systems fall into these independently selectable categories, and
# every combination is published. Independent rather than a cumulative ladder on purpose:
# "LLM-based systems but not the plain open-source ones" is a real question, and a ladder
# cannot express it. Mirrors `SYSTEM_CATEGORIES` in tabarena/evaluation/entrants.py, whose
# `pool_key` builds the folder segments below.
SYSTEM_CATEGORY_LABELS = {
    "open": "📊 Open-source systems",
    "llm": "🤖 Systems with LLMs",
    "api": "🔒 Closed-source API systems",
}

# The tag that puts a system in each category; None is the untagged (plain open-source) group.
SYSTEM_CATEGORY_TAGS = {
    "open": None,
    "llm": "with-llm",
    "api": "closed-source-api",
}

SYSTEM_CATEGORY_NOTES = {
    "open": "Systems you can inspect and run yourself, such as AutoGluon.",
    "llm": "Systems with an LLM in the loop, including agents.",
    "api": "Systems behind a remote API whose internals we cannot inspect.",
}

# Shown on a category that has no entrant yet, which is rendered as a disabled toggle.
CATEGORY_COMING_SOON = "Coming soon: waiting for a submission"

# Folder segment for the pool where no system competes.
MODELS_ONLY_KEY = "models"


def entrants_key(categories: Iterable[str]) -> str:
    """Folder segment for a set of selected categories, in `SYSTEM_CATEGORY_LABELS` order.

    Order-independent, so ticking the boxes in any order lands on the same artifacts.
    Mirrors `pool_key` in tabarena/evaluation/entrants.py.
    """
    selected = set(categories or ())
    ordered = [key for key in SYSTEM_CATEGORY_LABELS if key in selected]
    return "_".join(ordered) if ordered else MODELS_ONLY_KEY


def widest_entrants_key() -> str:
    """The pool where every category competes; the one to read totals from."""
    return entrants_key(SYSTEM_CATEGORY_LABELS)


def entrants_categories(key: str) -> list[str]:
    """The selected category keys encoded in a folder segment."""
    return [] if key == MODELS_ONLY_KEY else [k for k in SYSTEM_CATEGORY_LABELS if k in key.split("_")]


def entrants_name(key: str) -> str:
    """Human-readable name for a pool, used in figure labels."""
    selected = entrants_categories(key)
    if not selected:
        return "Models only"
    return "Models + " + ", ".join(SYSTEM_CATEGORY_LABELS[k] for k in selected)


def entrants_note(key: str) -> str:
    """One line describing who competes in a pool."""
    selected = entrants_categories(key)
    if not selected:
        return "Individual models only, each run under TabArena's shared tuning protocol."
    return "Also competing: " + " ".join(SYSTEM_CATEGORY_NOTES[k] for k in selected)


@lru_cache(maxsize=None)
def available_categories(data_root: str) -> frozenset[str]:
    """Which system categories actually have an entrant in the published artifacts.

    Read from the widest pool's leaderboard, so a category nobody has submitted to yet can be
    shown as a disabled toggle instead of a setting that silently changes nothing.
    """
    path = Path(data_root) / Subset(entrants=widest_entrants_key()).rel_path / "website_leaderboard.csv"
    if not path.exists():
        return frozenset()
    df = load_leaderboard_csv(str(path.resolve()))
    if "MethodClass" not in df.columns:
        return frozenset()
    systems = df[df["MethodClass"] == "system"]
    if systems.empty:
        return frozenset()
    # An empty Tags cell round-trips through CSV as NaN, and `str(nan)` is the string "nan",
    # so an untagged system would read as tagged. Test for null rather than truthiness.
    tag_sets = [
        set() if pd.isna(v) else {t for t in str(v).split(";") if t}
        for v in systems.get("Tags", pd.Series(dtype=object))
    ]
    found = set()
    for key, tag in SYSTEM_CATEGORY_TAGS.items():
        if tag is None:
            if any(not tags for tags in tag_sets):
                found.add(key)
        elif any(tag in tags for tags in tag_sets):
            found.add(key)
    return frozenset(found)


# axis -> {value: human label}. Insertion order = display order; first = default.
TASK_LABELS = {
    "all": "All Tasks",
    "classification": "Classification",
    "regression": "Regression",
    "binary": "Binary",
    "multiclass": "Multiclass",
}
DATASET_LABELS = {
    "all": "All Datasets",
    "small": "Small",
    "medium": "Medium",
}

# Short labels used as column headers in the cross-subset overview.
TASK_SHORT = {
    "all": "Overall",
    "classification": "Class.",
    "regression": "Regr.",
    "binary": "Binary",
    "multiclass": "Multi.",
}
DATASET_SHORT = {
    "small": "Small",
    "medium": "Medium",
}

# What each choice means, shown as a hover tooltip on the chip (see `main.taStampTitles`).
# Every selector in the control card carries one, so nothing has to be guessed from a label.
TASK_NOTES = {
    "all": "Every task type: binary and multiclass classification plus regression.",
    "classification": "Classification only, binary and multiclass together.",
    "regression": "Regression tasks only, scored with RMSE.",
    "binary": "Binary classification only, scored with ROC AUC.",
    "multiclass": "Multiclass classification only, scored with log-loss.",
}

DATASET_NOTES = {
    "all": "Every curated dataset, whatever its size.",
    "small": "Datasets with at most 10,000 training rows.",
    "medium": "Datasets with between 10,001 and 100,000 training rows.",
}

# What each row of the control card selects, hovered on the caption at its left. Keyed by the
# class on the row, because a CSS ::before caption cannot carry a title of its own; the title
# goes on the row and a chip's own tooltip still wins over it (see `main.taStampTitles`).
AXIS_NOTES = {
    "ta-row-entrants": (
        "Who is scored together. Each combination is evaluated separately, so switching re-rates "
        "everyone rather than hiding rows: Elo is pairwise over the entrants and Improvability is "
        "the gap to the best of them."
    ),
    "axis-care": (
        "What to optimise for. Reorders the figures and picks the time axis the Pareto front is "
        "plotted against."
    ),
    "axis-metric": (
        "Which headline metric the page leads with. The second figure stays pinned to the other "
        "one, so both are always on the page."
    ),
    "axis-tasks": "Restrict the leaderboard to one task type.",
    "axis-datasets": "Restrict the leaderboard to one dataset-size bucket.",
    "ta-row-protocol": "How the reported numbers were computed.",
}

PROTOCOL_NOTES = {
    "imputed": (
        "Include methods that cannot run on every dataset. Their missing results are imputed "
        "with a default RandomForest, which counts against them for not covering the benchmark."
    ),
    "lite": (
        "Score each experiment on one split (first fold, first repeat) instead of all repeats. "
        "Cheaper and less reliable, but usually a good proxy."
    ),
}

DATASET_SIZE_NOTE = {
    "small": "Small datasets have at most 10,000 training rows.",
    "medium": "Medium datasets have between 10,001 and 100,000 training rows.",
    "tabpfn": (
        "TabPFNv2-compatible datasets contain at most 10,000 samples, "
        "500 features, and 10 classes."
    ),
}


@dataclass(frozen=True)
class Subset:
    """One cell of the leaderboard grid (entrants x imputation x splits x tasks x datasets).

    ``rel_path`` mirrors ``get_website_folder_name`` in
    ``tabarena/evaluation/subset_grid.py`` segment for segment: the path *is* the subset's
    identity on both sides, so changing the layout means changing both.
    """

    entrants: str = "models"  # `entrants_key(...)` of the selected categories
    imputation: str = "yes"  # "yes" | "no"
    splits: str = "all"  # "all" | "lite"
    tasks: str = "all"  # see TASK_LABELS
    datasets: str = "all"  # see DATASET_LABELS

    @property
    def rel_path(self) -> str:
        return (
            f"entrants_{self.entrants}/"
            f"imputation_{self.imputation}/"
            f"splits_{self.splits}/"
            f"tasks_{self.tasks}/"
            f"datasets_{self.datasets}"
        )


#: Family name the artifacts used before systems became their own entrant class. Artifacts
#: generated then are still served: every BeyondArena subset, and any TabArena subset not yet
#: regenerated. Without this the rows keep a family nothing maps a colour or pill to and render
#: grey, so the name is normalized on read and the rest of the app only ever sees "System".
_LEGACY_FAMILY_NAMES = {"Reference Pipeline": Constants.system}


@lru_cache(maxsize=None)
def load_leaderboard_csv(path: str) -> pd.DataFrame:
    """Read a ``website_leaderboard.csv`` (cached; files are tiny and immutable)."""
    df = pd.read_csv(path)
    df = df.rename(columns={"1#": "#"})
    if "TypeName" in df.columns:
        df["TypeName"] = df["TypeName"].replace(_LEGACY_FAMILY_NAMES)
    return df


VARIANT_RE = re.compile(r"\((default|tuned \+ ensembled|tuned)\)")


def parse_model(model: str) -> tuple[str, str, str | None]:
    """Split a Model cell into (base name, variant, url).

    A cell looks like ``[TabFM (default)](https://…)``, optionally followed by an
    ``[X% IMPUTED]`` tag. Used both for display (``views.py``) and for the JSON
    records the API returns (``api.py``), so the two cannot disagree.
    """
    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 unzip_png(base_dir: Path, img_name: str) -> str:
    """Return the path to ``base_dir/img_name``.png, unzipping the ``.png.zip`` on first access."""
    base = Path(base_dir) / img_name
    img_path = base.with_suffix(".png")
    if img_path.exists():
        return str(img_path)
    with zipfile.ZipFile(base.with_suffix(".png.zip"), "r") as zipf:
        zipf.extractall(img_path.parent)
    return str(img_path)


@dataclass
class LBContainer:
    """Loads the artifacts for a single subset under a given data root."""

    data_root: Path
    subset: Subset
    name: str
    n_datasets: int | None = None
    blurb: str | None = None
    base_path: Path = field(init=False)

    def __post_init__(self) -> None:
        self.base_path = Path(self.data_root) / self.subset.rel_path
        for fname in self._listdir():
            match = re.match(r"n_datasets_(.+)", fname)
            if match:
                self.n_datasets = match.group(1)
                break

    def _listdir(self) -> list[str]:
        try:
            return [p.name for p in self.base_path.iterdir()]
        except FileNotFoundError:
            return []

    def load_df(self) -> pd.DataFrame:
        return load_leaderboard_csv(str((self.base_path / "website_leaderboard.csv").resolve())).copy()

    def has_image(self, img_name: str) -> bool:
        """Whether this subset ships a static ``img_name`` figure.

        TabArena subsets ship interactive explorers only, so this is False for them; the
        BeyondArena subsets and any pre-explorer artifacts still carry PNGs.
        """
        base = self.base_path / img_name
        return base.with_suffix(".png").exists() or base.with_suffix(".png.zip").exists()

    def image_path(self, img_name: str) -> str:
        """Return the path to ``img_name``.png, unzipping it on first access."""
        return unzip_png(self.base_path, img_name)

    def html_content(self, name: str) -> str | None:
        """Return the inline content of ``name``.html (a self-contained
        interactive plot generated by the tabarena artifact pipeline), or
        ``None`` when the subset's data predates these artifacts — callers
        fall back to the static PNG then.
        """
        path = self.base_path / f"{name}.html"
        try:
            return path.read_text(encoding="utf-8")
        except FileNotFoundError:
            return None


def subset_name(subset: Subset) -> str:
    """Human-readable name for a subset, used in figure labels."""
    impute = "with imputation" if subset.imputation == "yes" else "no imputation"
    split = "all repeats" if subset.splits == "all" else "Lite"
    return (
        f"{entrants_name(subset.entrants)} | {TASK_LABELS[subset.tasks]} "
        f"| {DATASET_LABELS[subset.datasets]} | {split} | {impute}"
    )


def subset_blurb(subset: Subset, n_datasets: int | None) -> str:
    """One-line description of the subset shown above its figures."""
    datasets_name = DATASET_LABELS[subset.datasets].lower()
    blurb = (
        f"Leaderboard for {n_datasets} datasets "
        f"({datasets_name}, {TASK_LABELS[subset.tasks].lower()}) "
    )
    if subset.splits == "lite":
        blurb += "for one split (1st fold, 1st repeat) "
    blurb += "including all "
    if subset.imputation == "yes":
        blurb += "(imputed) "
    blurb += "models."

    # Which entrants competed decides every number above, so it is said here too.
    entrant_note = entrants_note(subset.entrants)
    if entrant_note:
        blurb += f"<br>{entrant_note}"
    note = DATASET_SIZE_NOTE.get(subset.datasets)
    if note:
        blurb += f"<br>{note}"
    return blurb


# --------------------------------------------------------------------------- #
# BeyondArena subsets
#
# BeyondArena diverges from TabArena: there is no imputation/splits/tasks/datasets
# grid. Instead a single axis of subset dimensions (split regime, size bucket,
# feature dimensionality/type) is surfaced as one tab bar, and every leaderboard
# is always computed on the recommended `core` protocol (`["core", <dim>]`; the
# "full" subset is `core` with no extra filter). The artifacts are produced by
# scripts/run_generate_beyondarena_website_artifacts.py in the tabarena repo, whose
# `BEYOND_SUBSETS` keys must match the labels below.
# --------------------------------------------------------------------------- #

# label -> human name. Insertion order = tab-bar order; first = default. Groups are
# only used to draw section separators in the tab bar / copy.
BEYOND_SUBSET_LABELS = {
    "full": "Full",
    "random": "IID",
    "temporal": "Temporal",
    "grouped": "Grouped",
    "tiny": "Tiny",
    "small": "Small",
    "medium": "Medium",
    "large": "Large",
    "low-dim": "Low-dim",
    "high-dim": "High-dim",
    "text": "Text",
    "high-cardinality": "High-cardinality",
}

# One-line description shown above each subset's figures. Kept in sync with the
# BeyondArena subset predicates (see BeyondArenaContext.SUBSET_PREDICATES).
BEYOND_SUBSET_NOTE = {
    "full": "All BeyondArena datasets, on the recommended core protocol.",
    "random": "IID (randomly split) tasks only.",
    "temporal": "Temporally split tasks only: train on the past, test on the future.",
    "grouped": "Group-wise split tasks only, with disjoint groups between train and test.",
    "tiny": "Tiny datasets contain at most 1,000 training rows.",
    "small": "Small datasets contain between 1,001 and 10,000 training rows.",
    "medium": "Medium datasets contain between 10,001 and 100,000 training rows.",
    "large": "Large datasets contain between 100,001 and 1,000,000 training rows.",
    "low-dim": "Low-dimensional datasets have at most 100 columns after preprocessing.",
    "high-dim": "High-dimensional datasets have more than 100 columns after preprocessing.",
    "text": "Datasets that contain one or more text columns.",
    "high-cardinality": "Datasets that contain one or more high-cardinality categorical columns.",
}


@dataclass(frozen=True)
class BeyondSubset:
    """One cell of the BeyondArena leaderboard — a single subset dimension, always on core."""

    subset: str = "full"  # see BEYOND_SUBSET_LABELS

    @property
    def rel_path(self) -> str:
        return f"subsets/{self.subset}"


def beyond_subset_name(subset: BeyondSubset) -> str:
    """Human-readable name for a BeyondArena subset, used in figure labels."""
    return f"{BEYOND_SUBSET_LABELS[subset.subset]} · core"


def beyond_subset_blurb(subset: BeyondSubset, n_datasets: int | None) -> str:
    """One-line description of a BeyondArena subset shown above its figures."""
    human = BEYOND_SUBSET_LABELS[subset.subset].lower()
    blurb = (
        f"Leaderboard for {n_datasets} BeyondArena datasets ({human}), evaluated on the "
        "recommended <b>core</b> protocol."
    )
    note = BEYOND_SUBSET_NOTE.get(subset.subset)
    if note:
        blurb += f"<br>{note}"
    return blurb