app.py

from __future__ import annotations

import base64
import json
import os
import re
from fnmatch import fnmatch
from io import BytesIO
from pathlib import Path
from urllib.parse import quote

import gradio as gr
import pandas as pd
from huggingface_hub import HfApi, hf_hub_download

from theme import POVISLE_THEME
from views import (
    render_examples_tab,
    render_leaderboard_category_tab,
    render_leaderboard_filters,
    render_leaderboard_task_tab,
)

EXCLUDED_MODELS = {"LLaVA-v1.6-Vicuna-13B-HF"}
MODEL_TYPE_LABELS = {
    "open": "Open-weight",
    "proprietary": "Proprietary",
    "random": "Random",
}
APP_CSS = """
:root {
  --povisle-bg: #f8f7f6;
  --povisle-surface: #ffffff;
  --povisle-soft: #efedeb;
  --povisle-soft-red: #fff1f2;
  --povisle-soft-blue: #edf3f9;
  --povisle-ink: #181416;
  --povisle-muted: #5f585b;
  --povisle-faint: #8b8588;
  --povisle-line: #e3dfdc;
  --povisle-accent: #d7263d;
  --povisle-blue: #2c5f8a;
  --povisle-green: #2a7a4e;
  --povisle-shadow: 0 2px 16px rgba(39, 33, 31, 0.07);
  --povisle-shadow-lg: 0 8px 40px rgba(39, 33, 31, 0.11);
  --povisle-serif: "Source Serif 4", Georgia, serif;
  --povisle-sans: "DM Sans", system-ui, -apple-system, BlinkMacSystemFont, "Segoe UI", sans-serif;
  --povisle-mono: "JetBrains Mono", ui-monospace, SFMono-Regular, Menlo, monospace;
}
body {
  color-scheme: light;
  background: var(--povisle-bg) !important;
  color: var(--povisle-ink) !important;
  font-family: var(--povisle-sans) !important;
  line-height: 1.7;
}
.gradio-container {
  max-width: 1480px !important;
  margin: 0 auto !important;
  padding: 0 14px 42px !important;
  background: transparent !important;
  color: var(--povisle-ink) !important;
  font-family: var(--povisle-sans) !important;
}
.contain {
  gap: 0 !important;
}
footer {
  display: none !important;
}
.povisle-hero {
  background: transparent;
  border: 0;
  border-radius: 0;
  box-shadow: none;
  margin: 0 auto 34px;
  max-width: 1120px;
  padding: 84px 28px 48px;
  text-align: center;
}
.povisle-kicker {
  align-items: center;
  background: var(--povisle-soft-red);
  border: 1px solid rgba(215, 38, 61, 0.22);
  border-radius: 20px;
  color: var(--povisle-accent);
  display: inline-flex;
  font-size: 0.78rem;
  font-weight: 700;
  gap: 6px;
  letter-spacing: 0.04em;
  line-height: 1;
  margin: 0 0 26px;
  padding: 7px 14px;
  text-transform: uppercase;
}
.povisle-hero h1 {
  color: var(--povisle-ink) !important;
  display: block !important;
  font-family: var(--povisle-serif) !important;
  font-size: clamp(2.65rem, 4.5vw, 4.35rem) !important;
  font-weight: 600 !important;
  letter-spacing: 0 !important;
  line-height: 1.02 !important;
  margin: 0 auto 12px !important;
  max-width: 1200px !important;
}
.povisle-title-accent {
  color: var(--povisle-accent) !important;
}
.povisle-hero em {
  color: var(--povisle-ink);
  font-style: italic;
}
.povisle-hero p {
  color: var(--povisle-muted);
  font-size: 0.98rem;
  line-height: 1.8;
  margin: 26px auto 0;
  max-width: 720px;
}
.povisle-authors {
  color: var(--povisle-muted);
  font-size: 0.9rem;
  line-height: 1.65;
  margin: 24px auto 0;
  max-width: 760px;
}
.povisle-authors strong {
  color: var(--povisle-ink);
  font-weight: 500;
}
.povisle-links-row {
  display: flex;
  flex-wrap: wrap;
  gap: 10px;
  justify-content: center;
  margin-top: 30px;
}
.povisle-hero a.povisle-btn,
.povisle-hero a.povisle-btn:hover,
.povisle-hero a.povisle-btn:focus,
.povisle-hero a.povisle-btn:visited {
  text-decoration: none !important;
}
.povisle-btn {
  align-items: center;
  border: 1px solid transparent;
  border-radius: 6px;
  display: inline-flex;
  font-size: 0.84rem;
  font-weight: 500;
  gap: 7px;
  line-height: 1;
  padding: 10px 18px;
  text-decoration: none;
  transition: all 0.18s ease;
}
.povisle-btn svg {
  flex-shrink: 0;
  height: 15px;
  width: 15px;
}
.povisle-btn svg.arxiv-icon {
  height: 22.5px;
  width: 22.5px;
}
.povisle-btn-primary {
  background: var(--povisle-accent);
  color: #ffffff !important;
}
.povisle-btn-primary:hover {
  background: #b91c33;
  box-shadow: var(--povisle-shadow);
  transform: translateY(-1px);
}
.povisle-btn-dataset {
  background: var(--povisle-soft-blue);
  border-color: rgba(44, 95, 138, 0.2);
  color: var(--povisle-blue) !important;
}
.povisle-hero a.povisle-btn-dataset,
.povisle-hero a.povisle-btn-dataset:visited {
  color: var(--povisle-blue) !important;
}
.povisle-btn-dataset:hover {
  background: #dbe9f5;
  transform: translateY(-1px);
}
.povisle-btn-outline {
  background: var(--povisle-surface);
  border-color: var(--povisle-line);
  color: var(--povisle-ink) !important;
}
.povisle-hero a.povisle-btn-outline,
.povisle-hero a.povisle-btn-outline:visited {
  color: var(--povisle-ink) !important;
}
.povisle-btn-outline:hover {
  border-color: var(--povisle-ink);
  box-shadow: var(--povisle-shadow);
  transform: translateY(-1px);
}
.povisle-stat-row {
  display: grid;
  gap: 16px;
  grid-template-columns: repeat(auto-fit, minmax(150px, 1fr));
  margin: 30px auto 0;
  max-width: 760px;
}
.povisle-stat {
  background: var(--povisle-bg);
  border: 1px solid var(--povisle-line);
  border-radius: 10px;
  padding: 18px 14px;
  text-align: center;
}
.povisle-stat strong {
  color: var(--povisle-accent) !important;
  display: block;
  font-family: var(--povisle-serif);
  font-size: 1.7rem;
  font-weight: 600;
  line-height: 1.1;
}
.povisle-stat span {
  color: var(--povisle-faint);
  display: block;
  font-size: 0.78rem;
  font-weight: 500;
  margin-top: 5px;
}
.tab-nav {
  justify-content: center !important;
  margin: 0 auto !important;
  max-width: 860px !important;
}
.tab-nav button {
  color: var(--povisle-muted) !important;
  font-family: var(--povisle-sans) !important;
  font-size: 0.88rem !important;
  font-weight: 600 !important;
}
.tabitem {
  padding-top: 24px !important;
}
#leaderboard-section {
  padding-top: 0 !important;
}
#leaderboard-search {
  margin-bottom: 14px !important;
}
#leaderboard-filter-row {
  margin-bottom: 14px !important;
}
#category-view-select {
  margin-bottom: 14px !important;
}
.leaderboard-split-tabs {
  margin-top: 0 !important;
}
#leaderboard-table,
.leaderboard-table {
  margin-top: 0 !important;
}
.form,
.block {
  border-color: var(--povisle-line) !important;
  border-radius: 8px !important;
  box-shadow: none !important;
}
.wrap.default {
  background: transparent !important;
}
#leaderboard-filter-row {
  background: var(--povisle-surface) !important;
  border: 1px solid var(--povisle-line) !important;
  border-radius: 8px !important;
  gap: 0 !important;
  overflow: hidden !important;
}
#leaderboard-filter-row > *,
#split-filter,
#type-filter,
#size-filter,
#family-filter {
  background: var(--povisle-surface) !important;
  border-color: transparent !important;
  border-radius: 0 !important;
  box-shadow: none !important;
}
#split-filter {
  border-radius: 8px 0 0 8px !important;
}
#type-filter {
  border-left: 0 !important;
  border-radius: 8px 0 0 8px !important;
}
#family-filter {
  border-radius: 0 8px 8px 0 !important;
}
#type-filter,
#size-filter,
#family-filter {
  border-left: 1px dashed #ddd7d3 !important;
}
#type-filter {
  border-left: 0 !important;
}
#leaderboard-filter-row .form,
#leaderboard-filter-row .block {
  background: transparent !important;
  border: 0 !important;
  border-radius: 0 !important;
}
#leaderboard-filter-row [data-testid="token"],
#leaderboard-filter-row .token,
#leaderboard-filter-row .selected-token,
#leaderboard-filter-row .multiselect-token,
#leaderboard-filter-row div:has(> button[aria-label^="Remove"]),
#leaderboard-filter-row div:has(> button[aria-label^="remove"]),
#leaderboard-filter-row button[aria-label^="Remove"],
#leaderboard-filter-row button[aria-label^="remove"] {
  align-items: center !important;
  background: #efedeb !important;
  border: 1px solid #ddd7d3 !important;
  border-radius: 999px !important;
  color: var(--povisle-ink) !important;
  display: inline-flex !important;
  font-size: 0.82rem !important;
  font-weight: 500 !important;
  gap: 6px !important;
  line-height: 1 !important;
  min-height: 28px !important;
  padding: 5px 10px !important;
}
#leaderboard-filter-row [data-testid="token"] button,
#leaderboard-filter-row .token button,
#leaderboard-filter-row .selected-token button,
#leaderboard-filter-row .multiselect-token button {
  background: transparent !important;
  border: 0 !important;
  color: var(--povisle-muted) !important;
  min-height: auto !important;
  padding: 0 0 0 3px !important;
}
#leaderboard-filter-row [data-testid="token"] svg,
#leaderboard-filter-row .token svg,
#leaderboard-filter-row .selected-token svg,
#leaderboard-filter-row .multiselect-token svg {
  height: 12px !important;
  width: 12px !important;
}
#leaderboard-search,
#category-view-select,
#leaderboard-table,
.leaderboard-table,
#examples-table {
  border-radius: 10px !important;
}
#leaderboard-table,
.leaderboard-table,
#examples-table {
  box-shadow: var(--povisle-shadow);
}
#leaderboard-table table th,
.leaderboard-table table th,
#examples-table table th {
  background: var(--povisle-soft) !important;
  color: var(--povisle-ink) !important;
  font-size: 0.7rem !important;
  font-weight: 700 !important;
  line-height: 1.2 !important;
  padding: 8px 10px !important;
  white-space: normal !important;
}
#leaderboard-table table td,
.leaderboard-table table td,
#examples-table table td {
  border-color: var(--povisle-line) !important;
  color: var(--povisle-muted) !important;
}
#leaderboard-table table th,
.leaderboard-table table th,
#leaderboard-table table td,
.leaderboard-table table td {
  min-width: 8.75rem;
}
#leaderboard-table table th:nth-child(1),
.leaderboard-table table th:nth-child(1),
#leaderboard-table table td:nth-child(1),
.leaderboard-table table td:nth-child(1) {
  min-width: 7rem;
}
#leaderboard-table table th:nth-child(2),
.leaderboard-table table th:nth-child(2),
#leaderboard-table table td:nth-child(2),
.leaderboard-table table td:nth-child(2) {
  min-width: 15rem;
}
#leaderboard-table table th:nth-child(3),
.leaderboard-table table th:nth-child(3),
#leaderboard-table table td:nth-child(3),
.leaderboard-table table td:nth-child(3) {
  min-width: 7rem;
}
#leaderboard-table table th:nth-child(4),
.leaderboard-table table th:nth-child(4),
#leaderboard-table table td:nth-child(4),
.leaderboard-table table td:nth-child(4) {
  min-width: 8rem;
}
#leaderboard-table table tbody tr:nth-child(even) td,
.leaderboard-table table tbody tr:nth-child(even) td,
#examples-table table tbody tr:nth-child(even) td {
  background: #fbfaf9;
}
#leaderboard-table table tbody tr:hover td,
.leaderboard-table table tbody tr:hover td,
#examples-table table tbody tr:hover td {
  background: #fff1f2 !important;
}
.povisle-section-label {
  color: var(--povisle-accent) !important;
  font-size: 0.72rem;
  font-weight: 700;
  letter-spacing: 0.1em;
  margin-bottom: 12px;
  text-transform: uppercase;
}
.povisle-dataset-card .povisle-section-label,
.povisle-leaderboard-heading .povisle-section-label,
.povisle-examples-heading .povisle-section-label {
  color: var(--povisle-accent) !important;
}
.povisle-section-title {
  color: var(--povisle-ink);
  font-family: var(--povisle-serif);
  font-size: clamp(1.3rem, 2.5vw, 1.7rem);
  font-weight: 600;
  margin: 0 0 18px;
}
.povisle-dataset-card {
  background: transparent;
  border-color: transparent;
  border-radius: 0;
  box-shadow: none;
  margin: 30px auto 8px;
  max-width: 1440px;
  padding: 12px 0 4px;
}
.povisle-dataset-card p {
  color: var(--povisle-muted);
  font-size: 0.97rem;
  line-height: 1.8;
  margin: 0 0 14px;
  max-width: 100%;
}
.povisle-dataset-card p:last-child {
  margin-bottom: 0;
}
.povisle-about-grid {
  align-items: start;
  display: grid;
  gap: 28px;
  grid-template-columns: 1fr;
  margin-top: 26px;
}
.povisle-about-copy {
  margin-top: 22px;
}
.povisle-about-grid .povisle-about-copy {
  margin-top: 0;
}
.povisle-category-layout {
  align-items: start;
  display: grid;
  gap: 24px;
  grid-template-columns: repeat(2, minmax(0, 1fr));
}
.povisle-task-layout {
  align-items: start;
  display: grid;
  gap: 24px;
  grid-template-columns: repeat(2, minmax(0, 1fr));
}
.povisle-creation-layout {
  align-items: start;
  display: grid;
  gap: 24px;
  grid-template-columns: repeat(2, minmax(0, 1fr));
}
.povisle-creation-figure {
  background: var(--povisle-surface);
  border: 1px solid var(--povisle-line);
  border-radius: 8px;
  margin: 0;
  overflow: hidden;
  padding: 12px;
}
.povisle-creation-figure img {
  display: block;
  height: auto;
  width: 100%;
}
.povisle-category-intro {
  margin-bottom: 8px !important;
}
.povisle-about-copy h3 {
  color: var(--povisle-ink);
  font-family: var(--povisle-serif);
  font-size: 1.18rem;
  font-weight: 600;
  line-height: 1.25;
  margin: 0 0 10px;
}
.povisle-angled-sunburst {
  background: var(--povisle-surface) !important;
  border: 1px solid var(--povisle-line) !important;
  border-radius: 8px !important;
  box-shadow: none !important;
  margin: 0 !important;
  overflow: hidden !important;
  padding: 14px !important;
}
.povisle-angled-sunburst img {
  display: block;
  height: clamp(260px, 36vw, 520px);
  margin: 0 auto;
  max-width: 720px;
  object-fit: contain;
  width: 100%;
}
.povisle-examples-heading {
  margin: 38px auto 18px;
  max-width: 1440px;
}
.povisle-leaderboard-heading {
  margin: 26px auto 18px;
  max-width: 1440px;
}
.povisle-leaderboard-heading .povisle-section-title {
  margin-bottom: 10px;
}
.povisle-leaderboard-heading p {
  color: var(--povisle-muted);
  font-size: 0.97rem;
  line-height: 1.7;
  margin: 0;
  max-width: 100%;
}
.povisle-examples-heading p {
  color: var(--povisle-muted);
  font-size: 0.97rem;
  line-height: 1.8;
  margin: 0;
}
.povisle-acknowledgement {
  border-top: 1px solid var(--povisle-line);
  margin: 46px auto 0;
  max-width: 1440px;
  padding: 24px 0 0;
}
.povisle-acknowledgement .povisle-section-title {
  font-size: 1.15rem;
  margin-bottom: 8px;
}
.povisle-acknowledgement p {
  color: var(--povisle-muted);
  font-size: 0.94rem;
  line-height: 1.7;
  margin: 0;
  max-width: 100%;
}
.povisle-citation {
  border-top: 1px solid var(--povisle-line);
  margin: 30px auto 0;
  max-width: 1440px;
  padding: 24px 0 0;
}
.povisle-citation .povisle-section-title {
  font-size: 1.15rem;
  margin-bottom: 8px;
}
.povisle-citation pre {
  background: var(--povisle-surface);
  border: 1px solid var(--povisle-line);
  border-radius: 8px;
  color: var(--povisle-ink);
  font-family: var(--povisle-mono);
  font-size: 0.82rem;
  line-height: 1.55;
  margin: 0;
  overflow-x: auto;
  padding: 14px;
  white-space: pre;
}
#examples-filter-row {
  background: var(--povisle-surface) !important;
  border: 1px solid var(--povisle-line) !important;
  border-radius: 10px !important;
  gap: 0 !important;
  margin: 0 0 16px !important;
  overflow: hidden !important;
  padding: 0 !important;
  width: 100% !important;
}
#examples-filter-row > * {
  background: transparent !important;
  position: relative !important;
}
#examples-filter-row > * + *::before {
  background: #ddd7d3;
  bottom: 14px;
  content: "";
  left: 0;
  position: absolute;
  top: 14px;
  width: 1px;
  z-index: 1;
}
#examples-filter-row .block,
#examples-filter-row .form,
#examples-filter-row .wrap,
#examples-filter-row .wrap-inner,
#examples-filter-row .input-container {
  background: transparent !important;
  border: 0 !important;
  border-radius: 0 !important;
  box-shadow: none !important;
}
#examples-filter-row input,
#examples-filter-row textarea {
  background: transparent !important;
}
#examples-table {
  overflow-x: auto !important;
}
#examples-table table {
  table-layout: fixed !important;
  min-width: 1530px !important;
}
#examples-table table th,
#examples-table table td,
#examples-table [role="columnheader"],
#examples-table [role="gridcell"] {
  line-height: 1.35 !important;
  padding: 6px 8px !important;
}
#examples-table table td,
#examples-table [role="gridcell"] {
  font-size: 0.64rem !important;
  line-height: 1.3 !important;
}
#examples-table table th,
#examples-table [role="columnheader"] {
  font-size: 0.66rem !important;
  line-height: 1.15 !important;
  padding: 7px 8px !important;
}
#examples-table table td *,
#examples-table table th *,
#examples-table [role="gridcell"] *,
#examples-table [role="columnheader"] * {
  font-size: inherit !important;
  line-height: inherit !important;
}
#examples-table table td,
#examples-table table th,
#examples-table [role="gridcell"],
#examples-table [role="columnheader"] {
  min-width: 0 !important;
  overflow-wrap: normal !important;
  white-space: normal !important;
  word-break: normal !important;
}
#examples-table table td *,
#examples-table table th *,
#examples-table [role="gridcell"] *,
#examples-table [role="columnheader"] * {
  overflow-wrap: break-word !important;
  white-space: normal !important;
  word-break: normal !important;
}
#examples-table table td:nth-child(1),
#examples-table table th:nth-child(1) {
  width: 90px !important;
}
#examples-table table td:nth-child(2),
#examples-table table th:nth-child(2) {
  width: 128px !important;
}
#examples-table table td:nth-child(3),
#examples-table table th:nth-child(3) {
  width: 130px !important;
}
#examples-table table td:nth-child(4),
#examples-table table th:nth-child(4) {
  width: 110px !important;
  text-align: center;
  vertical-align: middle;
}
#examples-table table td:nth-child(5),
#examples-table table th:nth-child(5) {
  width: 300px !important;
}
#examples-table table td:nth-child(6),
#examples-table table th:nth-child(6) {
  width: 230px !important;
}
#examples-table table td:nth-child(6),
#examples-table [role="gridcell"]:nth-child(6) {
  white-space: pre-line !important;
}
#examples-table table td:nth-child(7),
#examples-table table th:nth-child(7) {
  width: 170px !important;
}
#examples-table table td:nth-child(8),
#examples-table table th:nth-child(8) {
  width: 190px !important;
}
#examples-table table td:nth-child(9),
#examples-table table th:nth-child(9) {
  width: 90px !important;
}
#examples-table table td:nth-child(10),
#examples-table table th:nth-child(10) {
  width: 100px !important;
}
#examples-table table td:nth-child(4) img.example-image {
  display: block;
  margin: 0 auto;
  max-height: 6rem;
  max-width: 6rem;
  object-fit: contain;
}
#leaderboard-section,
#examples-section {
  background: transparent !important;
  border: 0 !important;
  box-shadow: none !important;
}
.povisle-category-list {
  color: var(--povisle-muted);
  font-size: 0.88rem;
  line-height: 1.45;
  margin: 4px 0 0;
  padding-left: 1.1rem;
}
.povisle-category-list li + li {
  margin-top: 8px;
}
.povisle-category-list strong {
  color: var(--povisle-ink);
  font-weight: 720;
}
.povisle-category-note {
  margin-top: 18px !important;
}
.povisle-task-list {
  color: var(--povisle-muted);
  font-size: 0.9rem;
  line-height: 1.55;
  margin: 0;
  padding-left: 1.1rem;
}
.povisle-task-list li + li {
  margin-top: 8px;
}
.povisle-task-list strong {
  color: var(--povisle-ink);
  font-weight: 720;
}
.povisle-task-chart {
  background: var(--povisle-surface);
  border: 1px solid var(--povisle-line);
  border-radius: 8px;
  padding: 16px;
}
.povisle-task-chart-legend {
  display: flex;
  flex-wrap: wrap;
  gap: 10px 14px;
  margin-bottom: 16px;
}
.povisle-task-chart-legend span {
  align-items: center;
  color: var(--povisle-muted);
  display: inline-flex;
  font-size: 0.78rem;
  gap: 6px;
  line-height: 1.2;
}
.povisle-task-chart-legend span::before {
  background: var(--segment-color);
  border-radius: 999px;
  content: "";
  display: inline-block;
  height: 8px;
  width: 8px;
}
.povisle-task-split + .povisle-task-split {
  margin-top: 18px;
}
.povisle-task-split-header {
  align-items: baseline;
  display: flex;
  justify-content: space-between;
  gap: 12px;
  margin-bottom: 8px;
}
.povisle-task-split-header span {
  color: var(--povisle-muted);
  font-size: 0.82rem;
  line-height: 1.25;
}
.povisle-task-split-header strong {
  color: var(--povisle-ink);
  font-family: var(--povisle-mono);
  font-size: 0.82rem;
  font-weight: 650;
  text-align: right;
}
.povisle-task-chart-track {
  background: var(--povisle-soft);
  border-radius: 999px;
  display: flex;
  height: 24px;
  overflow: hidden;
}
.povisle-task-chart-segment {
  align-items: center;
  background: var(--segment-color);
  color: #ffffff;
  display: flex;
  font-family: var(--povisle-mono);
  font-size: 0.72rem;
  font-weight: 650;
  height: 100%;
  justify-content: center;
  line-height: 1;
  min-width: 28px;
}
.povisle-task-chart-segment + .povisle-task-chart-segment {
  border-left: 1px solid rgba(255, 255, 255, 0.75);
}
.block label span,
.block .label-wrap span {
  color: #3c4149 !important;
  font-weight: 680 !important;
}
@media (max-width: 760px) {
  .gradio-container {
    padding: 0 14px 32px !important;
  }
  .povisle-hero {
    padding: 48px 18px 38px;
  }
  .povisle-hero h1 {
    font-size: clamp(1.95rem, 9.5vw, 2.75rem) !important;
  }
  .povisle-links-row {
    align-items: stretch;
    flex-direction: column;
  }
  .povisle-btn {
    justify-content: center;
  }
  .povisle-dataset-card {
    padding: 8px 0 4px;
  }
  .povisle-about-grid {
    grid-template-columns: 1fr;
  }
  .povisle-category-layout {
    grid-template-columns: 1fr;
  }
  .povisle-task-layout {
    grid-template-columns: 1fr;
  }
  .povisle-creation-layout {
    grid-template-columns: 1fr;
  }
}

/* Leaderboard layout reset after adding nested split tabs. */
#leaderboard-section,
#leaderboard-section > *,
#leaderboard-section .tabitem,
#leaderboard-view-tabs,
#leaderboard-view-tabs > *,
#leaderboard-view-tabs .tabitem,
.leaderboard-split-tabs,
.leaderboard-split-tabs > *,
.leaderboard-split-tabs .tabitem {
  background: transparent !important;
}
#leaderboard-section {
  display: block !important;
  background: var(--povisle-bg) !important;
  margin-top: 0 !important;
  padding-top: 0 !important;
}
.tabitem:has(#leaderboard-section) {
  padding-top: 10px !important;
}
#leaderboard-section .form,
#leaderboard-section .block {
  box-shadow: none !important;
}
#leaderboard-search-row,
#leaderboard-filter-row,
#leaderboard-category-row {
  background: var(--povisle-surface) !important;
  border: 1px solid var(--povisle-line) !important;
  border-radius: 8px !important;
  box-shadow: none !important;
  filter: none !important;
  margin: 0 0 16px !important;
  overflow: hidden !important;
}
#leaderboard-search,
#category-view-select {
  background: transparent !important;
  border: 0 !important;
  border-radius: 0 !important;
  box-shadow: none !important;
  filter: none !important;
  margin: 0 !important;
  overflow: visible !important;
}
#leaderboard-search textarea,
#leaderboard-search input {
  background: var(--povisle-surface) !important;
}
#leaderboard-search-row,
#leaderboard-search-row *,
#leaderboard-category-row,
#leaderboard-category-row * {
  box-shadow: none !important;
  filter: none !important;
}
#leaderboard-search,
#leaderboard-search *,
#category-view-select,
#category-view-select * {
  box-shadow: none !important;
  filter: none !important;
}
#leaderboard-search,
#leaderboard-search .block,
#leaderboard-search .form,
#leaderboard-search .wrap,
#leaderboard-search .wrap-inner,
#leaderboard-search .input-container,
#leaderboard-search textarea,
#leaderboard-search input,
#category-view-select,
#category-view-select .block,
#category-view-select .form,
#category-view-select .wrap,
#category-view-select .wrap-inner,
#category-view-select .input-container {
  background: var(--povisle-surface) !important;
}
#leaderboard-search .block,
#leaderboard-search .form,
#leaderboard-search .wrap,
#leaderboard-search .wrap-inner,
#leaderboard-search .input-container,
#category-view-select .block,
#category-view-select .form,
#category-view-select .wrap,
#category-view-select .wrap-inner,
#category-view-select .input-container {
  border: 0 !important;
  border-radius: 0 !important;
}
#leaderboard-category-row {
  background: transparent !important;
  border: 0 !important;
  border-radius: 0 !important;
  padding: 12px 0 !important;
  overflow: visible !important;
}
#leaderboard-category-row > *,
#leaderboard-category-row #category-view-select,
#leaderboard-category-row #category-view-select .block,
#leaderboard-category-row #category-view-select .form,
#leaderboard-category-row #category-view-select .wrap,
#leaderboard-category-row #category-view-select .wrap-inner,
#leaderboard-category-row #category-view-select .input-container,
#leaderboard-category-row #category-view-select [role="radiogroup"] {
  background: transparent !important;
  border: 0 !important;
  box-shadow: none !important;
}
#category-view-select .wrap,
#category-view-select .wrap-inner,
#category-view-select .input-container,
#category-view-select [role="radiogroup"] {
  align-items: center !important;
  display: flex !important;
  flex-wrap: wrap !important;
  gap: 8px !important;
}
#category-view-select label {
  align-items: center !important;
  background: var(--povisle-soft) !important;
  border: 1px solid var(--povisle-line) !important;
  border-radius: 999px !important;
  color: var(--povisle-muted) !important;
  cursor: pointer !important;
  display: inline-flex !important;
  font-size: 0.86rem !important;
  font-weight: 600 !important;
  line-height: 1 !important;
  margin: 0 !important;
  max-width: 100% !important;
  min-height: 34px !important;
  padding: 9px 13px !important;
  transition: background 0.16s ease, border-color 0.16s ease, color 0.16s ease;
}
#category-view-select label:hover {
  background: #fff1f2 !important;
  border-color: #f2b8c0 !important;
  color: var(--povisle-ink) !important;
}
#category-view-select label:has(input:checked) {
  background: var(--povisle-accent) !important;
  border-color: var(--povisle-accent) !important;
  color: #ffffff !important;
}
#category-view-select label:has(input:focus-visible) {
  outline: 2px solid rgba(215, 38, 61, 0.26) !important;
  outline-offset: 2px !important;
}
#category-view-select input[type="radio"] {
  height: 1px !important;
  margin: 0 !important;
  opacity: 0 !important;
  position: absolute !important;
  width: 1px !important;
}
#category-view-select label span {
  color: inherit !important;
  white-space: nowrap !important;
}
#leaderboard-filter-row {
  gap: 0 !important;
  overflow: hidden !important;
}
#leaderboard-filter-row > * {
  background: var(--povisle-surface) !important;
  position: relative !important;
}
.povisle-filter-field,
.leaderboard-filter-field {
  background: var(--povisle-surface) !important;
  gap: 0 !important;
  padding: 14px 14px 12px !important;
}
.povisle-filter-field .block,
.povisle-filter-field .form,
.povisle-filter-field .wrap,
.povisle-filter-field .wrap-inner,
.povisle-filter-field .input-container,
.leaderboard-filter-field .block,
.leaderboard-filter-field .form,
.leaderboard-filter-field .wrap,
.leaderboard-filter-field .wrap-inner,
.leaderboard-filter-field .input-container {
  background: transparent !important;
  border: 0 !important;
  box-shadow: none !important;
}
#leaderboard-filter-row > * + *::before {
  background: #ddd7d3;
  bottom: 18px;
  content: "";
  left: 0;
  position: absolute;
  top: 18px;
  width: 1px;
}
#type-filter,
#size-filter,
#family-filter {
  background: var(--povisle-surface) !important;
  border: 0 !important;
  border-radius: 0 !important;
  box-shadow: none !important;
}
#size-filter,
#family-filter {
  border-left: 1px dashed #ddd7d3 !important;
}
#leaderboard-filter-row label,
#leaderboard-filter-row label span,
#leaderboard-filter-row .label-wrap,
#leaderboard-filter-row .label-wrap span,
#leaderboard-filter-row [data-testid="block-label"],
#leaderboard-filter-row [data-testid="block-label"] span {
  color: var(--povisle-ink) !important;
  opacity: 1 !important;
}
#leaderboard-filter-row .label-wrap span,
#leaderboard-filter-row [data-testid="block-label"] span {
  font-weight: 700 !important;
}
#leaderboard-filter-row input,
#leaderboard-filter-row textarea,
#leaderboard-filter-row [role="combobox"],
#leaderboard-filter-row [data-testid="token"],
#leaderboard-filter-row .token,
#leaderboard-filter-row .selected-token,
#leaderboard-filter-row .multiselect-token {
  color: var(--povisle-ink) !important;
}
#leaderboard-view-tabs {
  border: 0 !important;
  box-shadow: none !important;
  margin: 8px 0 0 !important;
  padding: 0 !important;
}
#leaderboard-view-tabs > .tab-nav {
  background: transparent !important;
  border-bottom: 1px solid var(--povisle-line) !important;
  justify-content: flex-start !important;
  margin: 0 0 16px !important;
  max-width: none !important;
  padding: 0 !important;
}
#leaderboard-view-tabs > .tab-nav button {
  font-size: 0.94rem !important;
  font-weight: 700 !important;
  padding: 10px 14px !important;
}
#leaderboard-view-tabs > .tabitem {
  border: 0 !important;
  padding: 0 !important;
}
.leaderboard-split-tabs {
  border: 0 !important;
  box-shadow: none !important;
  margin: 4px 0 0 !important;
  padding: 0 !important;
}
.leaderboard-split-tabs .tab-nav {
  background: transparent !important;
  border-bottom: 1px solid var(--povisle-line) !important;
  justify-content: flex-start !important;
  margin: 0 0 14px !important;
  max-width: none !important;
  padding: 0 !important;
}
.leaderboard-split-tabs .tab-nav button {
  font-size: 0.84rem !important;
  padding: 8px 12px !important;
}
.leaderboard-split-tabs .tabitem {
  border: 0 !important;
  padding: 0 !important;
}
.leaderboard-table {
  margin-top: 0 !important;
}
"""
APP_HEAD = """
<link rel="preconnect" href="https://fonts.googleapis.com">
<link rel="preconnect" href="https://fonts.gstatic.com" crossorigin>
<link href="https://fonts.googleapis.com/css2?family=Source+Serif+4:ital,opsz,wght@0,8..60,300;0,8..60,400;0,8..60,600;0,8..60,700;1,8..60,300;1,8..60,400&family=DM+Sans:ital,wght@0,300;0,400;0,500;0,600;1,300&family=JetBrains+Mono:wght@400;500&display=swap" rel="stylesheet">
"""
HERO_HTML = """
<section class="povisle-hero">
  <h1>
    <em>Jako Tako</em> or Fluent? Presenting <span class="povisle-title-accent">PoVisLE</span>
  </h1>
  <p>
    We introduce PoVisLE, a monocultural vision-language benchmark for Polish designed to evaluate
    culturally grounded multimodal understanding under a grounded evaluation paradigm, where language
    is interpreted in interaction with visual context. The dataset contains 1,117 images and 2,366
    manually annotated VQA pairs.
  </p>
  <div class="povisle-authors">
    <strong>Anna Ko&#322;os</strong> &middot; <strong>Grzegorz Statkiewicz</strong> &middot; <strong>Karolina Seweryn</strong> &middot;
    <strong>Katarzyna Kowol</strong> &middot; <strong>Karolina Piosek</strong> &middot; <strong>Wojciech Kusa</strong><br>
    NASK National Research Institute, Warsaw, Poland
  </div>
  <div class="povisle-links-row">
    <a href="https://huggingface.co/collections/NASK-PIB/povisle" class="povisle-btn povisle-btn-dataset" target="_blank">
      <svg width="15" height="15" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" aria-hidden="true"><ellipse cx="12" cy="5" rx="9" ry="3"/><path d="M21 12c0 1.66-4 3-9 3s-9-1.34-9-3"/><path d="M3 5v14c0 1.66 4 3 9 3s9-1.34 9-3V5"/></svg>
      Dataset (Validation)
    </a>
    <a href="https://github.com/NASK-NLP/PoVisLE" class="povisle-btn povisle-btn-outline" target="_blank">
      <svg width="15" height="15" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" aria-hidden="true"><path d="M9 19c-5 1.5-5-2.5-7-3m14 6v-3.87a3.37 3.37 0 0 0-.94-2.61c3.14-.35 6.44-1.54 6.44-7A5.44 5.44 0 0 0 20 4.77 5.07 5.07 0 0 0 19.91 1S18.73.65 16 2.48a13.38 13.38 0 0 0-7 0C6.27.65 5.09 1 5.09 1A5.07 5.07 0 0 0 5 4.77a5.44 5.44 0 0 0-1.5 3.78c0 5.42 3.3 6.61 6.44 7A3.37 3.37 0 0 0 9 18.13V22"/></svg>
      Code
    </a>
    <a href="https://arxiv.org/abs/2608.07763" class="povisle-btn povisle-btn-outline" target="_blank">
      <svg class="arxiv-icon" width="23" height="23" viewBox="0 0 448 512" aria-hidden="true"><path fill="currentColor" d="m 119.65,351.996 c -5.84961,0.13477 -11.1943,-3.29883 -13.502,-8.67578 -2.19629,-5.27051 -0.61914,-8.9668 4.19727,-15.8652 7.05469,-10.3799 78.8242,-96.5625 78.8242,-96.5625 l -15.8828,-14.8633 c -13.3809,-13.3779 -13.9561,-31.333 -1.50977,-43.7754 l 18.4922,-17.6133 -51.5977,-63.377 c -4.00586,-4.26758 -6.48535,-11.7559 -4.24805,-17.1309 2.27832,-5.53613 7.69727,-9.12891 13.6836,-9.07031 3.83398,0.0957 7.43262,1.87402 9.83789,4.86133 l 61.3691,57.0566 94.5762,-90.0703 c 3.19434,-3.08398 7.44336,-4.83984 11.8828,-4.91016 1.60449,0.0039 3.19824,0.245117 4.73242,0.714844 5.7793,1.80566 10.249,6.41797 11.8711,12.252 1.2998,5.47363 -0.27637,11.2334 -4.18164,15.2832 l -83.0859,100.011996 14.8789,13.834 c 11.0957,10.0029 11.1543,27.3906 0.12695,37.4688 l -16.2949,15.6309 56.2559,66.4434 0.0742,0.0859 0.0664,0.0899 c 5.02734,6.53125 7.43164,11.5615 4.83984,17.9395 -3.13379,5.96875 -8.69727,10.29 -15.2559,11.8516 -0.67676,0.0908 -1.35938,0.13574 -2.04297,0.13671 l -0.004,-0.0117 c -4.42871,-0.27051 -8.61035,-2.13086 -11.7754,-5.24023 l -0.13086,-0.10743 -0.12304,-0.11132 -65.207,-59.127 -89.8965,86.2402 c 0,0 -5.33398,6.47754 -10.9707,6.61133 z m 178.104,-33.041 c 0.47852,-0.002 0.95703,-0.0332 1.43164,-0.0957 4.84277,-1.33301 8.95605,-4.54004 11.4316,-8.91016 1.44922,-3.5625 1.00293,-6.45703 -4.19727,-13.2148 l -56.0586,-66.2188 -26.375,25.3027 64.9551,58.9062 c 2.33887,2.41504 5.46484,3.91113 8.8125,4.2168 v 0.0137 z M 192.436,227.402 334.446,57.244 c 2.78711,-3.48926 4.5293,-6.97949 3.3418,-10.9121 -1.16113,-4.30273 -4.43457,-7.71777 -8.68555,-9.05664 -1.08105,-0.333984 -2.20508,-0.503906 -3.33594,-0.505859 -3.20801,0.07617 -6.26953,1.35645 -8.57617,3.58789 l -142.033,135.284999 c -11.1719,11.1719 -9.48242,26.0195 1.39258,36.8945 z"/></svg>
      Paper
    </a>
  </div>
  <div class="povisle-stat-row" aria-label="Dataset summary">
    <div class="povisle-stat"><strong>1,117</strong><span>images</span></div>
    <div class="povisle-stat"><strong>2,366</strong><span>questions</span></div>
    <div class="povisle-stat"><strong>7</strong><span>main categories</span></div>
    <div class="povisle-stat"><strong>3</strong><span>task formats</span></div>
  </div>
</section>
"""
DATASET_CREATION_IMAGE_URL = f"/gradio_api/file={quote(str((Path('public') / 'dataset.png').resolve()))}"
CATEGORY_SUNBURST_URL = f"/gradio_api/file={quote(str((Path('public') / 'categories_sunburst.png').resolve()))}"
DATASET_DESCRIPTION_HTML = f"""
<section class="povisle-dataset-card">
  <div class="povisle-section-label">Dataset</div>
  <h2 class="povisle-section-title">About the Benchmark</h2>
  <p>
    PoVisLE is a monocultural vision-language evaluation benchmark centered on Polish cultural
    and linguistic competence. It is designed for grounded evaluation: answers should depend on
    the interaction between the image and the question, rather than on text-only associations or
    surface-level entity recognition.

  </p>
  <div class="povisle-about-grid">
    <div class="povisle-about-copy">
      <h3>Tasks</h3>
      <div class="povisle-task-layout">
        <div>
          <p>
            During annotation, each VQA pair is assigned one of three task types. These
            are multiple-choice, binary yes/no, and open-ended questions. All questions
            are designed to require image understanding, and the answer should not be
            obtainable from textual knowledge alone without reference to the image.
          </p>
          <ul class="povisle-task-list">
            <li><strong>Multiple-choice questions</strong> include answer options appended below the question with letter labels.</li>
            <li><strong>Yes/no questions</strong> use the raw question as the prompt and are evaluated as binary answers.</li>
            <li><strong>Open-ended questions</strong> also use the raw question as the prompt, without appended answer options.</li>
          </ul>
        </div>
        <div class="povisle-task-chart" aria-label="Question distribution by task format and split">
          <div class="povisle-task-chart-legend" aria-hidden="true">
            <span style="--segment-color: #d7263d;">Multiple choice</span>
            <span style="--segment-color: #2c5f8a;">Yes/no</span>
            <span style="--segment-color: #2a7a4e;">Open-ended</span>
          </div>
          <div class="povisle-task-split">
            <div class="povisle-task-split-header">
              <span>Test split</span>
              <strong>1,960</strong>
            </div>
            <div class="povisle-task-chart-track" title="Test: 643 multiple-choice, 714 yes/no, 603 open-ended">
              <div class="povisle-task-chart-segment" style="--segment-color: #d7263d; width: 32.8%;" aria-label="Multiple choice: 643, 32.8%">643</div>
              <div class="povisle-task-chart-segment" style="--segment-color: #2c5f8a; width: 36.4%;" aria-label="Yes/no: 714, 36.4%">714</div>
              <div class="povisle-task-chart-segment" style="--segment-color: #2a7a4e; width: 30.8%;" aria-label="Open-ended: 603, 30.8%">603</div>
            </div>
          </div>
          <div class="povisle-task-split">
            <div class="povisle-task-split-header">
              <span>Validation split</span>
              <strong>406</strong>
            </div>
            <div class="povisle-task-chart-track" title="Validation: 212 multiple-choice, 154 yes/no, 40 open-ended">
              <div class="povisle-task-chart-segment" style="--segment-color: #d7263d; width: 52.2%;" aria-label="Multiple choice: 212, 52.2%">212</div>
              <div class="povisle-task-chart-segment" style="--segment-color: #2c5f8a; width: 37.9%;" aria-label="Yes/no: 154, 37.9%">154</div>
              <div class="povisle-task-chart-segment" style="--segment-color: #2a7a4e; width: 9.9%;" aria-label="Open-ended: 40, 9.9%">40</div>
            </div>
          </div>
        </div>
      </div>
    </div>
    <div class="povisle-about-copy">
      <h3>Categories</h3>
      <div class="povisle-category-layout">
        <div>
          <p class="povisle-category-intro">
            The taxonomy is adapted from PLCC and refined for the VQA setting. Grammar
            and vocabulary are merged into a unified Language category, while the hierarchy
            supports fine-grained diagnostics across cultural, linguistic, and visual
            reasoning domains. The dataset is organized into the following main categories.
          </p>
          <ul class="povisle-category-list">
            <li><strong>Art and Entertainment</strong> covers Polish and Poland-related artistic, media, and cultural references, including architecture, film, literature, music, paintings, sculpture, sport, and media.</li>
            <li><strong>Culture and Tradition</strong> covers shared customs, practices, symbols, cuisine, religion, traditions, pop culture, and regional or ethnic cultural variation.</li>
            <li><strong>Geography and Nature</strong> covers Polish physical, natural, urban, infrastructural, and socio-political spaces, including landscapes, landmarks, regions, and administrative entities.</li>
            <li><strong>History and Society</strong> covers historical and contemporary Polish social context, including the Middle Ages, World War II, post-war history, and current affairs.</li>
            <li><strong>Language</strong> covers visually grounded Polish linguistic phenomena, such as colloquial speech, slang, dialects, regionalisms, grammar, orthography, phraseology, rhetorical figures, and semantics.</li>
            <li><strong>Image Understanding</strong> focuses on direct recognition and interpretation of visual content in Polish or Poland-related contexts.</li>
            <li><strong>Visual Reasoning</strong> focuses on relationships, context, and inferred information within an image.</li>
          </ul>
          <p class="povisle-category-note">
            Image Understanding and Visual Reasoning form a smaller complementary subset
            focused on general multimodal skills while remaining embedded in Polish visual
            and linguistic contexts.
          </p>
        </div>
        <div class="povisle-angled-sunburst" aria-label="Static sunburst of categories and subcategories">
          <img src="{CATEGORY_SUNBURST_URL}" alt="PoVisLE category and subcategory distribution">
        </div>
      </div>
    </div>
  </div>
  <div class="povisle-about-copy">
    <h3>Dataset Creation</h3>
    <div class="povisle-creation-layout">
      <div>
        <p>
          PoVisLE was created through manual, template-free annotation. Annotators
          selected or reviewed images from Wikimedia Commons, other permissively
          available public sources, and personal collections contributed for research
          use. Each image was paired with one or more Polish VQA prompts and labeled
          with a task type, category, and subcategory.
        </p>
        <p>
          The dataset construction also included a Wikimedia-based augmentation stage
          to increase visual diversity and reduce selection bias. Candidate images
          were reviewed by annotators, and visually similar replacements were used
          only when the original question remained answerable from the new image.
        </p>
        <p>
          Quality assurance included cross-validation by a second annotator, metadata
          and license checks, regular team discussion, and supervision by an expert
          annotator. The annotation guidelines required questions to be visually
          grounded, unambiguous, linguistically natural, and suitable for deterministic
          evaluation.
        </p>
        <p>
          The benchmark is divided into a held-out test split for final evaluation and
          a public validation split. The validation split is intended mainly to
          illustrate the range of question types, so it is not sampled from the same
          distribution as the test set. In particular, it contains fewer open-ended
          questions, which are largely retained in the held-out test split.
        </p>
      </div>
      <div class="povisle-creation-figure">
        <img src="{DATASET_CREATION_IMAGE_URL}" alt="Overview of the PoVisLE dataset construction process">
      </div>
    </div>
  </div>
  <div class="povisle-about-copy">
    <h3>Evaluation Protocol</h3>
    <p>
      Models receive the image and a Polish prompt specifying the expected answer format, length,
      word order, and, where relevant, grammatical form. Macro accuracy is the main metric.
      Multiple-choice questions use circular evaluation: answer options are cyclically rotated, and
      a prediction is counted as correct only if the model selects the gold answer under every
      rotation. This reduces option-position bias while keeping evaluation efficient.
    </p>
    <p>
      Yes/no and open-ended questions are evaluated in a single pass. Yes/no questions require a
      binary answer in Polish. Open-ended predictions are compared against the gold answers, with
      correct diacritics required in all cases and correct capitalization required where relevant.
      For selected questions, multiple answer variants are accepted through predefined inclusion
      patterns developed through iterative human validation.
    </p>
  </div>
</section>
"""
LEADERBOARD_SPLITS = ("test", "validation")
RESULTS_REPO = os.getenv("RESULTS_REPO", "NASK-PIB/PoVisLE-results")
SOURCE_DATASET_REPO = os.getenv("SOURCE_DATASET_REPO", "NASK-PIB/PoVisLE")
SOURCE_DATASET_CONFIGS = [
    config.strip()
    for config in os.getenv("SOURCE_DATASET_CONFIGS", "mcq,open,yn").split(",")
    if config.strip()
]
EXAMPLES_DATASET_SPLITS = [
    split.strip()
    for split in os.getenv("EXAMPLES_DATASET_SPLITS", os.getenv("EXAMPLES_DATASET_SPLIT", "validation")).split(",")
    if split.strip()
]
SOURCE_DATASET_REVISION = os.getenv("SOURCE_DATASET_REVISION", "v1.2.0")
HF_TOKEN = os.getenv("HF_TOKEN")
CACHE_PATH = Path(os.getenv("HF_HOME", ".")).expanduser()
LOCAL_RESULTS_PATH = Path(os.getenv("LOCAL_RESULTS_PATH", CACHE_PATH / "PoVisLE-results")).expanduser()
PUBLIC_PATH = Path("public").expanduser()
PUBLIC_IMAGES_PATH = Path(os.getenv("PUBLIC_IMAGES_PATH", PUBLIC_PATH / "images")).expanduser()
LEADERBOARD_RELATIVE_PATH = Path("leaderboard")
BENCHMARK_VERSION = os.getenv("BENCHMARK_VERSION", "v1.2.0")
CATEGORY_DISPLAY_ORDER = [
    "Art & Entertainment",
    "Culture & Tradition",
    "Geography & Nature",
    "History & Society",
    "Language",
    "Image Understanding",
    "Visual Reasoning",
]
CATEGORY_ORDER_ALIASES = {
    "art entertainment": "Art & Entertainment",
    "art and entertainment": "Art & Entertainment",
    "culture tradition": "Culture & Tradition",
    "culture and tradition": "Culture & Tradition",
    "geography nature": "Geography & Nature",
    "geography and nature": "Geography & Nature",
    "history society": "History & Society",
    "history and society": "History & Society",
    "language": "Language",
    "image understanding": "Image Understanding",
    "visual reasoning": "Visual Reasoning",
}


def model_size_bucket(model_size: str | None) -> str:
    if not model_size:
        return "Unknown"

    normalized = str(model_size).strip().upper()
    if not normalized.endswith("B"):
        return "Unknown"

    try:
        size_value = float(normalized[:-1])
    except ValueError:
        return "Unknown"

    if size_value < 10:
        return "1-10B"
    if size_value < 20:
        return "10-20B"
    if size_value < 40:
        return "20-40B"
    if size_value <= 100:
        return "30-100B"
    return ">100B"


def result_file_patterns() -> list[str]:
    return [
        f"leaderboard/*/*/{BENCHMARK_VERSION}/test/circular_circular/results.json",
        f"leaderboard/*/*/{BENCHMARK_VERSION}/validation/circular_circular/results.json",
    ]


def matching_result_files(repo_files: list[str], patterns: list[str]) -> list[str]:
    return sorted(path for path in repo_files if any(fnmatch(path, pattern) for pattern in patterns))


def refresh_results_from_hub(local_results_path: Path) -> None:
    patterns = result_file_patterns()
    repo_files = HfApi(token=HF_TOKEN).list_repo_files(repo_id=RESULTS_REPO, repo_type="dataset")
    matched_files = matching_result_files(repo_files, patterns)
    if not matched_files:
        raise FileNotFoundError(
            f"No result files matching {patterns} were found in dataset {RESULTS_REPO}. "
            f"Check RESULTS_REPO, BENCHMARK_VERSION={BENCHMARK_VERSION}, and HF_TOKEN access."
        )

    local_results_path.mkdir(parents=True, exist_ok=True)
    for filename in matched_files:
        hf_hub_download(
            repo_id=RESULTS_REPO,
            filename=filename,
            repo_type="dataset",
            local_dir=str(local_results_path),
            etag_timeout=30,
            token=HF_TOKEN,
        )


def resolve_results_dir() -> Path:
    env_path = os.environ.get("VPLCC_RESULTS_DIR")
    if env_path:
        candidate = Path(env_path).expanduser()
        if candidate.exists():
            return candidate

    try:
        refresh_results_from_hub(LOCAL_RESULTS_PATH)
    except Exception as error:
        fallback = LOCAL_RESULTS_PATH / LEADERBOARD_RELATIVE_PATH
        if fallback.exists():
            return fallback
        raise FileNotFoundError(
            f"Could not download results dataset from {RESULTS_REPO} and no local cache was found. ({error})"
        ) from error

    leaderboard_path = LOCAL_RESULTS_PATH / LEADERBOARD_RELATIVE_PATH
    if leaderboard_path.exists():
        return leaderboard_path
    raise FileNotFoundError(f"Could not find leaderboard directory inside downloaded dataset: {leaderboard_path}")


def format_category_label(category_name: str) -> str:
    return str(category_name).replace("_", " ").title()


def make_subcategory_column_key(category_name: str, subcategory_name: str) -> str:
    return f"__subcategory__::{category_name}::{subcategory_name}"


def percentage(value: float) -> float:
    return round(float(value) * 100, 2)


def build_category_scores(metrics: dict) -> dict[str, float]:
    return {
        format_category_label(category_name): percentage(category_metrics["macro_accuracy"])
        for category_name, category_metrics in metrics.get("by_category", {}).items()
    }


def build_subcategory_scores(metrics: dict) -> tuple[dict[str, float], dict[str, list[tuple[str, str]]]]:
    subcategory_scores: dict[str, float] = {}
    subcategory_groups: dict[str, list[tuple[str, str]]] = {}

    for category_name, subcategory_metrics in metrics.get("by_category_and_subcategory", {}).items():
        category_label = format_category_label(category_name)
        group_columns: list[tuple[str, str]] = []

        for subcategory_name, details in subcategory_metrics.items():
            column_key = make_subcategory_column_key(category_label, subcategory_name)
            subcategory_label = format_category_label(subcategory_name)
            group_columns.append((column_key, subcategory_label))
            subcategory_scores[column_key] = percentage(details["macro_accuracy"])

        if group_columns:
            subcategory_groups[category_label] = group_columns

    return subcategory_scores, subcategory_groups


def build_task_scores(metrics: dict) -> dict[str, float]:
    task_scores = {}
    task_column_labels = {
        "mcq": "MCQ",
        "yn": "Y/N",
        "open": "Open",
    }
    for task_name, task_metrics in metrics.get("by_task", {}).items():
        normalized_task_name = normalize_result_task_name(task_name)
        column_label = task_column_labels.get(normalized_task_name)
        if column_label:
            task_scores[column_label] = percentage(task_metrics["accuracy"])
    return task_scores


def iter_results_paths(results_dir: Path, split: str):
    return sorted(results_dir.glob(f"*/*/{BENCHMARK_VERSION}/{split}/circular_circular/results.json"))


def normalize_result_task_name(task_name: object) -> str:
    normalized = str(task_name).strip().lower().replace("_", "-")
    if normalized in {"yes/no", "yes-no", "yesno"}:
        return "yn"
    return normalized


def results_path_priority(results_path: Path, payload: dict) -> tuple[int, int, int, int, str]:
    metadata = payload.get("run_metadata", {})
    metrics = payload.get("metrics", {})
    mode = str(metadata.get("evaluation_mode") or results_path.parent.name).lower().replace("-", "_")
    ablation_markers = ("no_image", "without_image", "no_question", "without_question")
    is_ablation = any(metadata.get(marker) for marker in ablation_markers) or any(
        marker in mode for marker in ablation_markers
    )
    full_input_priority = 0 if is_ablation else 1
    expected_tasks = {normalize_result_task_name(task_name) for task_name in SOURCE_DATASET_CONFIGS}
    result_tasks = {
        normalize_result_task_name(task_name)
        for task_name in (metadata.get("tasks") or metrics.get("by_task", {}).keys())
    }
    all_tasks_priority = 1 if expected_tasks and expected_tasks.issubset(result_tasks) else 0
    total_priority = int(metrics.get("overall", {}).get("total") or 0)
    default_mode_priority = 1 if mode == "circular_circular" else 0
    return full_input_priority, all_tasks_priority, total_priority, default_mode_priority, str(results_path)


def sort_category_columns(category_labels: set[str]) -> list[str]:
    priority = {label: index for index, label in enumerate(CATEGORY_DISPLAY_ORDER)}

    def category_sort_key(label: str) -> tuple[int, str]:
        normalized = re.sub(r"[^a-z0-9]+", " ", str(label).lower()).strip()
        canonical_label = CATEGORY_ORDER_ALIASES.get(normalized, label)
        return priority.get(canonical_label, len(CATEGORY_DISPLAY_ORDER)), label

    return sorted(category_labels, key=category_sort_key)


def safe_image_filename(task_id: object, index: object | None = None) -> str:
    raw_name = str(task_id)
    if index is not None:
        raw_name = f"{raw_name}-{index}"
    safe_name = re.sub(r"[^A-Za-z0-9_.-]+", "_", raw_name).strip("._")
    return f"{safe_name or 'image'}.png"


def public_image_path(task_id: object, index: object | None = None) -> Path:
    return PUBLIC_IMAGES_PATH.resolve() / safe_image_filename(task_id, index)


def public_file_url(path: Path) -> str:
    return f"/gradio_api/file={quote(str(path.resolve()))}"


def pil_image_from_value(image: object):
    if image is None:
        return None

    try:
        from PIL import Image
    except ImportError:
        return None

    image_bytes: bytes | None = None

    if isinstance(image, dict):
        raw_bytes = image.get("bytes")
        image_path = image.get("path")
        if raw_bytes:
            image_bytes = raw_bytes
        elif image_path:
            candidate = Path(image_path)
            if candidate.exists():
                image_bytes = candidate.read_bytes()
    elif isinstance(image, bytes):
        image_bytes = image
    elif isinstance(image, str):
        if image.startswith("data:image/") and "," in image:
            image = image.split(",", 1)[1]
        try:
            candidate = Path(image)
            if len(image) < 512 and candidate.exists():
                image_bytes = candidate.read_bytes()
        except OSError:
            image_bytes = None
        if image_bytes is None:
            try:
                image_bytes = base64.b64decode(image, validate=True)
            except Exception:
                image_bytes = None
    elif hasattr(image, "copy") and hasattr(image, "save"):
        return image.copy()

    if image_bytes is None:
        return None

    try:
        return Image.open(BytesIO(image_bytes)).copy()
    except Exception:
        return None


def save_image_thumbnail(image: object, output_path: Path, max_size: tuple[int, int] = (128, 128)) -> bool:
    if output_path.exists():
        return True

    pil_image = pil_image_from_value(image)
    if pil_image is None:
        return False

    thumbnail = pil_image.copy()
    thumbnail.thumbnail(max_size)
    if thumbnail.mode not in {"RGB", "RGBA"}:
        thumbnail = thumbnail.convert("RGB")

    output_path.parent.mkdir(parents=True, exist_ok=True)
    thumbnail.save(output_path, format="PNG")
    return True


def image_file_to_html(image_path: Path) -> str:
    if not image_path.exists():
        return ""

    return f'<img class="example-image" src="{public_file_url(image_path)}" alt="task image" loading="lazy">'


def format_choices(item: dict[str, object]) -> str:
    choices = []
    for label in "ABCDEFGH":
        value = item.get(label)
        if value is None or pd.isna(value):
            continue
        text = str(value).strip()
        if text:
            choices.append(f"{label}. {text}")
    return "\n".join(choices)


def format_example_value(value: object) -> object:
    if isinstance(value, (list, dict)):
        return json.dumps(value, ensure_ascii=False)
    if value is None or pd.isna(value):
        return ""
    return value


def load_examples_rows() -> list[dict[str, object]]:
    try:
        from datasets import load_dataset
    except ImportError as error:
        print(f"Could not load examples because datasets is not installed: {error}")
        return []

    rows: list[dict[str, object]] = []
    loaded_sources: set[tuple[str, str]] = set()
    for config_name in SOURCE_DATASET_CONFIGS:
        dataset = None
        selected_split = None
        errors: list[str] = []
        for split_name in EXAMPLES_DATASET_SPLITS:
            load_kwargs = {
                "path": SOURCE_DATASET_REPO,
                "name": config_name,
                "split": split_name,
                "token": HF_TOKEN,
            }
            if SOURCE_DATASET_REVISION:
                load_kwargs["revision"] = SOURCE_DATASET_REVISION

            try:
                dataset = load_dataset(**load_kwargs)
                selected_split = split_name
                break
            except Exception as error:
                errors.append(f"{split_name}: {error}")

        if dataset is None or selected_split is None:
            print(f"Could not load examples from {SOURCE_DATASET_REPO}/{config_name}. Tried {', '.join(errors)}.")
            continue

        loaded_sources.add((config_name, selected_split))
        for item in dataset:
            image_path = public_image_path(item["id"])
            image_html = image_file_to_html(image_path) if save_image_thumbnail(item.get("image"), image_path) else ""
            rows.append(
                {
                    "id": item["id"],
                    "image": image_html,
                    "task": item.get("task") or config_name,
                    "category": item.get("category"),
                    "subcategory": item.get("subcategory"),
                    "question": item.get("question"),
                    "choices": format_choices(item),
                    "answer": item.get("answer"),
                    "include": format_example_value(item.get("include")),
                    "check_casing": format_example_value(item.get("check_casing")),
                    "check_diacritics": format_example_value(item.get("check_diacritics")),
                }
            )

    if rows:
        sources = ", ".join(f"{config}/{split}" for config, split in sorted(loaded_sources))
        print(f"Loaded {len(rows)} examples from {SOURCE_DATASET_REPO}: {sources}.")
    else:
        print(f"No examples were loaded from {SOURCE_DATASET_REPO}.")
    return rows


def load_dashboard_data() -> tuple[pd.DataFrame, list[str], dict[str, list[tuple[str, str]]], pd.DataFrame, str]:
    results_dir = resolve_results_dir()
    leaderboard_rows = []
    latest_results_by_model: dict[tuple[str, str], tuple[Path, tuple[int, int, int, int, str]]] = {}
    category_labels: set[str] = set()
    all_subcategory_groups: dict[str, list[tuple[str, str]]] = {}

    for split in LEADERBOARD_SPLITS:
        for results_path in iter_results_paths(results_dir, split):
            payload = json.loads(results_path.read_text(encoding="utf-8"))
            metadata = payload["run_metadata"]
            metadata_split = metadata.get("split")
            if metadata_split and str(metadata_split).lower() != split:
                continue
            model_name = metadata["model_name"]
            if model_name in EXCLUDED_MODELS:
                continue
            priority = results_path_priority(results_path, payload)
            key = (split, model_name)
            if key not in latest_results_by_model or priority > latest_results_by_model[key][1]:
                latest_results_by_model[key] = (results_path, priority)

    for (split, model_name), (results_path, _) in sorted(latest_results_by_model.items()):
        payload = json.loads(results_path.read_text(encoding="utf-8"))
        metadata = payload["run_metadata"]
        metrics = payload["metrics"]
        category_scores = build_category_scores(metrics)
        subcategory_scores, subcategory_groups = build_subcategory_scores(metrics)
        task_scores = build_task_scores(metrics)
        overall_score = percentage(metrics["overall"]["macro_accuracy"])
        category_labels.update(category_scores)
        for category_label, columns in subcategory_groups.items():
            known_column_keys = {column_key for column_key, _ in all_subcategory_groups.get(category_label, [])}
            all_subcategory_groups.setdefault(category_label, [])
            for column_key, column_label in columns:
                if column_key not in known_column_keys:
                    all_subcategory_groups[category_label].append((column_key, column_label))
                    known_column_keys.add(column_key)
        leaderboard_rows.append(
            {
                "Split": split.title(),
                "Org": metadata.get("org", "unknown"),
                "Model": model_name,
                "Type": MODEL_TYPE_LABELS.get(metadata["model_type"], metadata["model_type"].title()),
                "Family": metadata.get("model_family", "unknown"),
                "Size": metadata.get("model_size", "unknown"),
                "Size range": model_size_bucket(metadata.get("model_size")),
                "Overall": overall_score,
                **task_scores,
                **category_scores,
                **subcategory_scores,
            }
        )

    category_columns = sort_category_columns(category_labels)
    leaderboard = pd.DataFrame(leaderboard_rows)
    for task_column in ["MCQ", "Y/N", "Open"]:
        if task_column not in leaderboard:
            leaderboard[task_column] = pd.NA
    for category_column in category_columns:
        if category_column not in leaderboard:
            leaderboard[category_column] = pd.NA
    for subcategory_columns in all_subcategory_groups.values():
        for column_key, _ in subcategory_columns:
            if column_key not in leaderboard:
                leaderboard[column_key] = pd.NA

    leaderboard = leaderboard.sort_values(by=["Split", "Overall", *category_columns], ascending=[True, False, *([False] * len(category_columns))])
    leaderboard.insert(0, "Rank", range(1, len(leaderboard) + 1))
    examples_df = pd.DataFrame(
        load_examples_rows(),
        columns=[
            "id",
            "image",
            "task",
            "category",
            "subcategory",
            "question",
            "choices",
            "answer",
            "include",
            "check_casing",
            "check_diacritics",
        ],
    )
    return leaderboard, category_columns, all_subcategory_groups, examples_df, str(results_dir)


PUBLIC_IMAGES_PATH.mkdir(parents=True, exist_ok=True)
gr.set_static_paths(paths=[PUBLIC_PATH.resolve(), PUBLIC_IMAGES_PATH.resolve()])

LEADERBOARD_DF, CATEGORY_COLUMNS, SUBCATEGORY_GROUPS, EXAMPLES_DF, RESULTS_DIR = load_dashboard_data()

with gr.Blocks(
    title="PoVisLE",
) as demo:
    gr.HTML(HERO_HTML)
    with gr.Group(elem_id="leaderboard-section"):
        gr.HTML(
            """
            <section class="povisle-leaderboard-heading">
              <div class="povisle-section-label">Results</div>
              <h2 class="povisle-section-title">Leaderboard</h2>
              <p>Compare model performance by category or task type across the test and validation splits. Overall and category scores report <strong>macro accuracy</strong>, while task columns report task-level accuracy.</p>
            </section>
            """
        )
        search_box, type_filter, size_filter, family_filter = render_leaderboard_filters(LEADERBOARD_DF)
        with gr.Tabs(elem_id="leaderboard-view-tabs"):
            with gr.Tab("By category"):
                render_leaderboard_category_tab(
                    LEADERBOARD_DF,
                    CATEGORY_COLUMNS,
                    SUBCATEGORY_GROUPS,
                    search_box,
                    type_filter,
                    size_filter,
                    family_filter,
                )
            with gr.Tab("By task"):
                render_leaderboard_task_tab(
                    LEADERBOARD_DF,
                    search_box,
                    type_filter,
                    size_filter,
                    family_filter,
                )
    gr.HTML(DATASET_DESCRIPTION_HTML)
    render_examples_tab(EXAMPLES_DF)
    gr.HTML(
        """
        <section class="povisle-acknowledgement">
          <h2 class="povisle-section-title">Acknowledgement</h2>
          <p>This work was supported by the Polish Ministry of Digital Affairs (subsidy no. 4/WII/DBI/2026). The computational resources were provided by the Polish high-performance computing infrastructure PLGrid (HPC Center: ACK Cyfronet AGH) under computational grant no. PLG/2026/019138.</p>
        </section>
        <section class="povisle-citation">
          <h2 class="povisle-section-title">Citation</h2>
          <pre>@article{kolos2026povisle,
  title   = {Jako Tako or Fluent? Presenting PoVisLE: A Polish Vision-Language Evaluation},
  author  = {Ko{\\l}os, Anna and Statkiewicz, Grzegorz and Seweryn, Karolina and Kowol, Katarzyna and Piosek, Karolina and Kusa, Wojciech},
  journal = {arXiv preprint},
  year    = {2026}
}</pre>
        </section>
        """
    )


if __name__ == "__main__":
    demo.launch(
        theme=POVISLE_THEME,
        css=APP_CSS,
        head=APP_HEAD,
    )