Spaces:
Running
Running
Update app.py
Browse files
app.py
CHANGED
|
@@ -1,401 +1,363 @@
|
|
| 1 |
# -*- coding: utf-8 -*-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 2 |
import gradio as gr
|
| 3 |
import pandas as pd
|
| 4 |
from apscheduler.schedulers.background import BackgroundScheduler
|
| 5 |
-
# Removed Hugging Face Hub imports as they are not needed for the simplified leaderboard
|
| 6 |
|
| 7 |
-
#
|
|
|
|
|
|
|
|
|
|
| 8 |
try:
|
| 9 |
-
# Assume these contain the *content* without excessive inline styling
|
| 10 |
from src.about import (
|
| 11 |
CITATION_BUTTON_LABEL,
|
| 12 |
CITATION_BUTTON_TEXT,
|
| 13 |
-
EVALUATION_QUEUE_TEXT,
|
| 14 |
INTRODUCTION_TEXT,
|
| 15 |
LLM_BENCHMARKS_TEXT,
|
| 16 |
-
TITLE,
|
| 17 |
)
|
| 18 |
-
|
| 19 |
try:
|
| 20 |
-
from src.display.css_html_js import custom_css
|
| 21 |
except ImportError:
|
| 22 |
-
print("Warning: src.display.css_html_js not found.
|
| 23 |
-
custom_css = ""
|
| 24 |
|
| 25 |
-
from src.envs import REPO_ID
|
| 26 |
-
from src.submission.submit import add_new_eval
|
| 27 |
-
print("Successfully imported from src module.")
|
| 28 |
-
# Option 2: Placeholder definitions (REMOVE IF USING OPTION 1)
|
| 29 |
except ImportError:
|
| 30 |
-
|
| 31 |
-
|
| 32 |
-
|
| 33 |
-
|
| 34 |
-
|
| 35 |
-
TITLE
|
| 36 |
-
|
| 37 |
-
|
| 38 |
-
|
| 39 |
-
|
| 40 |
-
|
| 41 |
-
|
| 42 |
-
|
| 43 |
-
|
| 44 |
-
|
| 45 |
-
|
| 46 |
-
|
| 47 |
-
|
| 48 |
-
|
| 49 |
-
|
| 50 |
-
|
| 51 |
-
|
| 52 |
-
|
| 53 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 54 |
data = [
|
| 55 |
-
{
|
| 56 |
-
|
| 57 |
-
|
| 58 |
-
|
| 59 |
-
|
| 60 |
-
|
| 61 |
-
|
| 62 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 63 |
]
|
| 64 |
master_df = pd.DataFrame(data)
|
|
|
|
| 65 |
CATEGORIES = ["Overall", "MLE-Lite", "Tabular", "NLP", "CV"]
|
| 66 |
DEFAULT_CATEGORY = "Overall"
|
| 67 |
category_to_column = {
|
| 68 |
-
"MLE-Lite": "MLE-Lite_Elo",
|
| 69 |
-
"
|
|
|
|
|
|
|
|
|
|
| 70 |
}
|
| 71 |
|
| 72 |
-
#
|
| 73 |
-
|
| 74 |
-
|
| 75 |
-
|
| 76 |
-
|
| 77 |
-
"""
|
| 78 |
-
score_column = category_to_column.get(category)
|
| 79 |
-
|
| 80 |
-
|
| 81 |
-
|
| 82 |
-
|
| 83 |
-
|
| 84 |
-
|
| 85 |
-
|
| 86 |
-
|
| 87 |
-
|
| 88 |
-
|
| 89 |
-
|
| 90 |
-
|
| 91 |
-
|
| 92 |
-
df.reset_index(drop=True, inplace=True)
|
| 93 |
-
df.insert(0, 'Rank', df.index + 1)
|
| 94 |
-
|
| 95 |
-
# Format Model Name as HTML Hyperlink - use a CSS class for styling
|
| 96 |
-
df['Model'] = df.apply(
|
| 97 |
-
lambda row: f"<a href='{row['url'] if pd.notna(row['url']) else '#'}' target='_blank' class='model-link'>{row['model_name']}</a>",
|
| 98 |
-
axis=1
|
| 99 |
)
|
| 100 |
|
| 101 |
-
|
| 102 |
-
|
| 103 |
-
|
| 104 |
-
|
| 105 |
-
|
| 106 |
-
|
| 107 |
-
# --- Mock/Placeholder functions/data for other tabs ---
|
| 108 |
-
print("Warning: Evaluation queue data fetching is disabled/mocked due to leaderboard changes.")
|
| 109 |
-
finished_eval_queue_df = pd.DataFrame(columns=["Model", "Status", "Requested", "Started"])
|
| 110 |
-
running_eval_queue_df = pd.DataFrame(columns=["Model", "Status", "Requested", "Started"])
|
| 111 |
-
pending_eval_queue_df = pd.DataFrame(columns=["Model", "Status", "Requested", "Started"])
|
| 112 |
-
EVAL_COLS = ["Model", "Status", "Requested", "Started"]
|
| 113 |
-
EVAL_TYPES = ["str", "str", "str", "str"]
|
| 114 |
-
|
| 115 |
-
# --- Keep restart function if relevant ---
|
| 116 |
-
def restart_space():
|
| 117 |
-
print(f"Attempting to restart space: {REPO_ID}")
|
| 118 |
-
# Replace with your actual space restart mechanism if needed
|
| 119 |
-
|
| 120 |
-
# --- Enhanced CSS Definition ---
|
| 121 |
-
# Define all styles here. Assumes TITLE has id="main-leaderboard-title"
|
| 122 |
-
# and INTRODUCTION_TEXT is wrapped in class="introduction-section" (or rendered by gr.Markdown).
|
| 123 |
-
|
| 124 |
-
enhanced_css = """
|
| 125 |
-
/* Base and Theme Overrides */
|
| 126 |
-
body {
|
| 127 |
-
font-family: -apple-system, BlinkMacSystemFont, 'Segoe UI', Roboto, Oxygen, Ubuntu, Cantarell, 'Open Sans', 'Helvetica Neue', sans-serif;
|
| 128 |
-
font-size: 1.3em; /* Base font size */
|
| 129 |
-
line-height: 1.6;
|
| 130 |
-
background-color: #f8f9fa; /* Light background */
|
| 131 |
-
color: #343a40; /* Default text color */
|
| 132 |
-
}
|
| 133 |
-
|
| 134 |
-
/* Container adjustments for better spacing */
|
| 135 |
-
.gradio-container {
|
| 136 |
-
max-width: 1200px !important; /* Limit max width */
|
| 137 |
-
margin: 0 auto !important; /* Center the container */
|
| 138 |
-
padding: 2rem !important; /* Add padding around the whole app */
|
| 139 |
-
}
|
| 140 |
-
|
| 141 |
-
/* --- Title Styling --- */
|
| 142 |
-
/* Targets the h1 tag with the specific ID from src/about.py */
|
| 143 |
-
#main-leaderboard-title {
|
| 144 |
-
font-size: 3.2em; /* Large title */
|
| 145 |
-
font-weight: 700; /* Bolder */
|
| 146 |
-
color: #212529; /* Darker color for title */
|
| 147 |
-
text-align: center; /* Ensure centering */
|
| 148 |
-
margin-bottom: 1.5rem; /* Space below title */
|
| 149 |
-
padding-bottom: 0.5rem; /* Space within the element */
|
| 150 |
-
border-bottom: 2px solid #dee2e6; /* Subtle underline */
|
| 151 |
-
}
|
| 152 |
|
| 153 |
-
|
| 154 |
-
|
| 155 |
-
|
| 156 |
-
|
| 157 |
-
|
| 158 |
-
color: #495057; /* Slightly lighter text color */
|
| 159 |
-
margin-bottom: 1rem; /* Space between paragraphs */
|
| 160 |
-
max-width: 900px; /* Limit width for readability */
|
| 161 |
-
margin-left: auto; /* Center the text block */
|
| 162 |
-
margin-right: auto; /* Center the text block */
|
| 163 |
-
text-align: center; /* Center align intro text */
|
| 164 |
-
}
|
| 165 |
-
.introduction-section, .introduction-wrapper {
|
| 166 |
-
font-family: 'Georgia',
|
| 167 |
-
font-size: 1.5em; !important; /* Slightly larger than base */
|
| 168 |
-
margin-bottom: 2.5rem; /* Space below the intro block */
|
| 169 |
-
}
|
| 170 |
|
|
|
|
|
|
|
|
|
|
| 171 |
|
| 172 |
-
|
| 173 |
-
|
| 174 |
-
|
| 175 |
-
|
| 176 |
-
|
| 177 |
-
|
| 178 |
-
|
| 179 |
-
|
| 180 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 181 |
}
|
| 182 |
-
.
|
| 183 |
-
font-size: 1.5em; /* Standard paragraph size */
|
| 184 |
margin-bottom: 1rem;
|
| 185 |
-
color: #495057;
|
| 186 |
}
|
| 187 |
-
|
| 188 |
-
|
| 189 |
-
|
| 190 |
-
|
| 191 |
}
|
| 192 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 193 |
font-size: 1.3em;
|
| 194 |
-
|
|
|
|
|
|
|
| 195 |
}
|
| 196 |
|
| 197 |
-
|
| 198 |
-
|
| 199 |
-
|
| 200 |
-
padding:
|
| 201 |
-
font-weight: 500;
|
| 202 |
}
|
| 203 |
|
| 204 |
-
|
| 205 |
-
|
| 206 |
-
|
| 207 |
-
|
| 208 |
-
|
| 209 |
-
|
|
|
|
|
|
|
| 210 |
}
|
|
|
|
|
|
|
| 211 |
#leaderboard-table th {
|
| 212 |
-
background-color: #e9ecef;
|
| 213 |
font-size: 1.3em;
|
| 214 |
-
font-weight: 500;
|
| 215 |
-
padding: 10px 12px; /* Header padding */
|
| 216 |
-
text-align: left;
|
| 217 |
-
color: #495057;
|
| 218 |
-
white-space: nowrap; /* Prevent header text wrapping */
|
| 219 |
}
|
| 220 |
#leaderboard-table td {
|
| 221 |
font-size: 1.1em;
|
| 222 |
-
padding: 8px 12px; /* Cell padding */
|
| 223 |
-
border-bottom: 1px solid #e9ecef; /* Horizontal lines */
|
| 224 |
-
vertical-align: middle; /* Center cell content vertically */
|
| 225 |
}
|
| 226 |
-
#leaderboard-table tr:nth-child(even) td {
|
| 227 |
-
font-size: 1.1em;
|
| 228 |
-
background-color: #f8f9fa; /* Zebra striping */
|
| 229 |
-
}
|
| 230 |
-
#leaderboard-table tr:hover td {
|
| 231 |
-
font-size: 1.1em;
|
| 232 |
-
background-color: #e2e6ea; /* Hover effect */
|
| 233 |
-
}
|
| 234 |
-
/* Style for the model links within the table */
|
| 235 |
#leaderboard-table .model-link {
|
| 236 |
-
color: #0056b3;
|
| 237 |
-
font-size: 1.1em;
|
| 238 |
font-weight: 500;
|
| 239 |
text-decoration: none;
|
| 240 |
}
|
| 241 |
#leaderboard-table .model-link:hover {
|
| 242 |
-
font-size: 1.1em;
|
| 243 |
text-decoration: underline;
|
| 244 |
-
color: #003d80;
|
| 245 |
-
}
|
| 246 |
-
|
| 247 |
-
/* --- Radio Button / Category Selector Styling --- */
|
| 248 |
-
.gradio-radio label span { /* Target the label text */
|
| 249 |
-
font-size: 1.3em !important;
|
| 250 |
-
font-weight: 500;
|
| 251 |
-
color: #343a40;
|
| 252 |
-
}
|
| 253 |
-
.gradio-radio fieldset { /* Adjust spacing around radio buttons */
|
| 254 |
-
margin-top: 0.5rem;
|
| 255 |
-
margin-bottom: 1.5rem;
|
| 256 |
-
}
|
| 257 |
-
.gradio-radio fieldset label { /* Style individual radio choices */
|
| 258 |
-
padding: 8px 12px !important;
|
| 259 |
-
}
|
| 260 |
-
|
| 261 |
-
|
| 262 |
-
/* --- Accordion Styling --- */
|
| 263 |
-
.gradio-accordion > button { /* Accordion header */
|
| 264 |
-
font-size: 1.2em !important;
|
| 265 |
-
font-weight: 600;
|
| 266 |
-
padding: 12px 15px !important;
|
| 267 |
-
background-color: #f1f3f5 !important;
|
| 268 |
-
border-bottom: 1px solid #dee2e6 !important;
|
| 269 |
-
}
|
| 270 |
-
.gradio-accordion > div { /* Accordion content area */
|
| 271 |
-
padding: 15px !important;
|
| 272 |
-
border: 1px solid #dee2e6 !important;
|
| 273 |
-
border-top: none !important;
|
| 274 |
}
|
| 275 |
-
|
| 276 |
-
/* --- Textbox/Button Styling (e.g., Citation) --- */
|
| 277 |
-
#citation-button textarea {
|
| 278 |
-
font-family: 'Courier New', Courier, monospace; /* Monospace for code/citation */
|
| 279 |
-
font-size: 0.95em !important;
|
| 280 |
-
background-color: #e9ecef;
|
| 281 |
-
color: #343a40;
|
| 282 |
-
}
|
| 283 |
-
#citation-button label span {
|
| 284 |
-
font-weight: 600;
|
| 285 |
-
}
|
| 286 |
-
|
| 287 |
-
|
| 288 |
-
/* ---- INTRODUCTION TEXT ---- */
|
| 289 |
-
.introduction-section {
|
| 290 |
-
font-size: 1.4rem !important;
|
| 291 |
-
line-height: 1.75;
|
| 292 |
-
color: #344054;
|
| 293 |
-
text-align: center;
|
| 294 |
-
max-width: 900px;
|
| 295 |
-
margin: 0 auto 3rem auto;
|
| 296 |
-
}
|
| 297 |
-
|
| 298 |
-
.introduction-section p {
|
| 299 |
-
margin-bottom: 1rem;
|
| 300 |
-
font-family: Georgia, serif;
|
| 301 |
-
font-size: 1.2em;
|
| 302 |
-
}
|
| 303 |
-
|
| 304 |
-
@media (max-width: 768px) {
|
| 305 |
-
.introduction-section {
|
| 306 |
-
font-size: 1.2rem !important;
|
| 307 |
-
}
|
| 308 |
-
}
|
| 309 |
-
|
| 310 |
-
|
| 311 |
"""
|
| 312 |
|
| 313 |
-
#
|
| 314 |
-
|
| 315 |
-
|
|
|
|
|
|
|
|
|
|
| 316 |
|
| 317 |
-
|
| 318 |
-
# Use a theme for base styling and apply custom CSS overrides
|
| 319 |
-
demo = gr.Blocks(css=final_css, theme=gr.themes.Soft(
|
| 320 |
-
# Optional: Customize theme variables if needed
|
| 321 |
-
# primary_hue=gr.themes.colors.blue,
|
| 322 |
-
# secondary_hue=gr.themes.colors.gray,
|
| 323 |
-
# neutral_hue=gr.themes.colors.cool_gray,
|
| 324 |
-
))
|
| 325 |
|
| 326 |
with demo:
|
| 327 |
-
#
|
| 328 |
gr.HTML(TITLE)
|
| 329 |
|
| 330 |
-
#
|
| 331 |
-
# Add a wrapper class for CSS targeting if the text itself doesn't have one
|
| 332 |
with gr.Row():
|
| 333 |
-
gr.Markdown(INTRODUCTION_TEXT, elem_classes="introduction-wrapper")
|
| 334 |
|
| 335 |
-
with gr.Tabs(elem_classes="tab-buttons")
|
| 336 |
-
|
|
|
|
| 337 |
with gr.Column():
|
| 338 |
-
# Use standard Markdown for the section header, CSS will style it
|
| 339 |
gr.Markdown("## Model Elo Rankings by Category", elem_classes="markdown-text")
|
|
|
|
| 340 |
category_selector = gr.Radio(
|
| 341 |
choices=CATEGORIES,
|
| 342 |
-
label="Select Category:",
|
| 343 |
value=DEFAULT_CATEGORY,
|
| 344 |
interactive=True,
|
| 345 |
-
elem_classes="gradio-radio"
|
| 346 |
)
|
|
|
|
| 347 |
leaderboard_df_component = gr.Dataframe(
|
| 348 |
value=update_leaderboard(DEFAULT_CATEGORY),
|
| 349 |
-
headers=["Rank", "Model", "Organizer", "License", "Elo
|
| 350 |
datatype=["number", "html", "str", "str", "number"],
|
| 351 |
interactive=False,
|
| 352 |
row_count=(len(master_df), "fixed"),
|
| 353 |
col_count=(5, "fixed"),
|
| 354 |
wrap=True,
|
| 355 |
-
elem_id="leaderboard-table"
|
| 356 |
)
|
| 357 |
-
category_selector.change(
|
| 358 |
-
fn=update_leaderboard,
|
| 359 |
-
inputs=category_selector,
|
| 360 |
-
outputs=leaderboard_df_component
|
| 361 |
-
)
|
| 362 |
-
|
| 363 |
-
with gr.TabItem("📝 About", elem_id="llm-benchmark-tab-about", id=1):
|
| 364 |
-
# Render LLM_BENCHMARKS_TEXT using Markdown, styled by CSS
|
| 365 |
-
gr.Markdown(LLM_BENCHMARKS_TEXT, elem_classes="markdown-text") # Apply standard markdown styling
|
| 366 |
|
| 367 |
-
|
| 368 |
-
# with gr.TabItem("🚀 Submit here! ", elem_id="llm-benchmark-tab-submit", id=2):
|
| 369 |
-
# # ... (Your submission form code - apply elem_classes="markdown-text" to gr.Markdown) ...
|
| 370 |
-
# pass # Placeholder
|
| 371 |
|
|
|
|
|
|
|
|
|
|
| 372 |
|
| 373 |
-
#
|
| 374 |
-
with gr.Accordion("📙 Citation", open=False, elem_classes="gradio-accordion"):
|
| 375 |
-
|
| 376 |
value=CITATION_BUTTON_TEXT,
|
| 377 |
label=CITATION_BUTTON_LABEL,
|
| 378 |
-
lines=8,
|
| 379 |
-
elem_id="citation-button",
|
| 380 |
show_copy_button=True,
|
| 381 |
)
|
| 382 |
|
| 383 |
-
#
|
|
|
|
|
|
|
| 384 |
if __name__ == "__main__":
|
| 385 |
try:
|
| 386 |
-
|
| 387 |
-
|
| 388 |
-
|
| 389 |
-
|
| 390 |
-
|
| 391 |
-
print("Scheduler started for space restart.")
|
| 392 |
-
else:
|
| 393 |
-
print("Warning: REPO_ID not set or is placeholder; space restart job not scheduled.")
|
| 394 |
else:
|
| 395 |
-
print("
|
| 396 |
-
except Exception as
|
| 397 |
-
print(f"
|
| 398 |
|
| 399 |
-
print("Launching Gradio
|
| 400 |
-
|
| 401 |
-
demo.launch()
|
|
|
|
| 1 |
# -*- coding: utf-8 -*-
|
| 2 |
+
"""Gradio frontend for the MLE‑Dojo leaderboard.
|
| 3 |
+
|
| 4 |
+
Changes made in this version
|
| 5 |
+
----------------------------
|
| 6 |
+
1. **Fixed CSS syntax errors** (missing semicolons, misplaced `!important`).
|
| 7 |
+
2. **Introduced a single, clean rule‑set** for the introduction block so the
|
| 8 |
+
font size, family and alignment are now reliably applied.
|
| 9 |
+
3. All new comments are in English for clarity.
|
| 10 |
+
"""
|
| 11 |
+
|
| 12 |
import gradio as gr
|
| 13 |
import pandas as pd
|
| 14 |
from apscheduler.schedulers.background import BackgroundScheduler
|
|
|
|
| 15 |
|
| 16 |
+
# ---------------------------------------------------------------------------
|
| 17 |
+
# Optional imports from the project package. If they fail we fall back to
|
| 18 |
+
# placeholders so the app still launches locally.
|
| 19 |
+
# ---------------------------------------------------------------------------
|
| 20 |
try:
|
|
|
|
| 21 |
from src.about import (
|
| 22 |
CITATION_BUTTON_LABEL,
|
| 23 |
CITATION_BUTTON_TEXT,
|
| 24 |
+
EVALUATION_QUEUE_TEXT, # still referenced in commented‑out submit tab
|
| 25 |
INTRODUCTION_TEXT,
|
| 26 |
LLM_BENCHMARKS_TEXT,
|
| 27 |
+
TITLE, # contains <h1 id="main-leaderboard-title">
|
| 28 |
)
|
| 29 |
+
|
| 30 |
try:
|
| 31 |
+
from src.display.css_html_js import custom_css # extra project CSS
|
| 32 |
except ImportError:
|
| 33 |
+
print("Warning: src.display.css_html_js not found. Using empty CSS.")
|
| 34 |
+
custom_css = ""
|
| 35 |
|
| 36 |
+
from src.envs import REPO_ID
|
| 37 |
+
from src.submission.submit import add_new_eval
|
|
|
|
|
|
|
| 38 |
except ImportError:
|
| 39 |
+
# -------- PLACEHOLDERS so the Space can still run --------
|
| 40 |
+
print("Warning: using placeholder values because src module imports failed.")
|
| 41 |
+
CITATION_BUTTON_LABEL = "Citation"
|
| 42 |
+
CITATION_BUTTON_TEXT = "Please cite us if you use this benchmark…"
|
| 43 |
+
EVALUATION_QUEUE_TEXT = "Current evaluation queue:"
|
| 44 |
+
TITLE = (
|
| 45 |
+
"<h1 id=\"main-leaderboard-title\" align=\"center\">🏆 MLE-Dojo "
|
| 46 |
+
"Benchmark Leaderboard (Placeholder)</h1>"
|
| 47 |
+
)
|
| 48 |
+
INTRODUCTION_TEXT = (
|
| 49 |
+
"<div class=\"introduction-section\">"
|
| 50 |
+
"<p>Welcome to the MLE‑Dojo Benchmark Leaderboard (placeholder).</p>"
|
| 51 |
+
"<p>Edit <code>src/about.py</code> to change this text.</p>"
|
| 52 |
+
"</div>"
|
| 53 |
+
)
|
| 54 |
+
LLM_BENCHMARKS_TEXT = (
|
| 55 |
+
"## About Section (placeholder)\nInformation about the benchmarks goes here."
|
| 56 |
+
)
|
| 57 |
+
custom_css = ""
|
| 58 |
+
REPO_ID = "your/space-id"
|
| 59 |
+
|
| 60 |
+
# Dummy function so the callback in the (commented) submit tab still works
|
| 61 |
+
def add_new_eval(*_):
|
| 62 |
+
return "Submission placeholder."
|
| 63 |
+
|
| 64 |
+
# ---------------------------------------------------------------------------
|
| 65 |
+
# Leaderboard data (static demo data for now)
|
| 66 |
+
# ---------------------------------------------------------------------------
|
| 67 |
+
|
| 68 |
data = [
|
| 69 |
+
{
|
| 70 |
+
"model_name": "gpt-4o-mini",
|
| 71 |
+
"url": "https://openai.com/index/gpt-4o-mini-advancing-cost-efficient-intelligence/",
|
| 72 |
+
"organizer": "OpenAI",
|
| 73 |
+
"license": "Proprietary",
|
| 74 |
+
"MLE-Lite_Elo": 753,
|
| 75 |
+
"Tabular_Elo": 839,
|
| 76 |
+
"NLP_Elo": 758,
|
| 77 |
+
"CV_Elo": 754,
|
| 78 |
+
"Overall": 778,
|
| 79 |
+
},
|
| 80 |
+
{
|
| 81 |
+
"model_name": "gpt-4o",
|
| 82 |
+
"url": "https://openai.com/index/hello-gpt-4o/",
|
| 83 |
+
"organizer": "OpenAI",
|
| 84 |
+
"license": "Proprietary",
|
| 85 |
+
"MLE-Lite_Elo": 830,
|
| 86 |
+
"Tabular_Elo": 861,
|
| 87 |
+
"NLP_Elo": 903,
|
| 88 |
+
"CV_Elo": 761,
|
| 89 |
+
"Overall": 841,
|
| 90 |
+
},
|
| 91 |
+
{
|
| 92 |
+
"model_name": "o3-mini",
|
| 93 |
+
"url": "https://openai.com/index/openai-o3-mini/",
|
| 94 |
+
"organizer": "OpenAI",
|
| 95 |
+
"license": "Proprietary",
|
| 96 |
+
"MLE-Lite_Elo": 1108,
|
| 97 |
+
"Tabular_Elo": 1019,
|
| 98 |
+
"NLP_Elo": 1056,
|
| 99 |
+
"CV_Elo": 1207,
|
| 100 |
+
"Overall": 1096,
|
| 101 |
+
},
|
| 102 |
+
{
|
| 103 |
+
"model_name": "deepseek-v3",
|
| 104 |
+
"url": "https://api-docs.deepseek.com/news/news1226",
|
| 105 |
+
"organizer": "DeepSeek",
|
| 106 |
+
"license": "DeepSeek",
|
| 107 |
+
"MLE-Lite_Elo": 1004,
|
| 108 |
+
"Tabular_Elo": 1015,
|
| 109 |
+
"NLP_Elo": 1028,
|
| 110 |
+
"CV_Elo": 1067,
|
| 111 |
+
"Overall": 1023,
|
| 112 |
+
},
|
| 113 |
+
{
|
| 114 |
+
"model_name": "deepseek-r1",
|
| 115 |
+
"url": "https://api-docs.deepseek.com/news/news250120",
|
| 116 |
+
"organizer": "DeepSeek",
|
| 117 |
+
"license": "DeepSeek",
|
| 118 |
+
"MLE-Lite_Elo": 1137,
|
| 119 |
+
"Tabular_Elo": 1053,
|
| 120 |
+
"NLP_Elo": 1103,
|
| 121 |
+
"CV_Elo": 1083,
|
| 122 |
+
"Overall": 1100,
|
| 123 |
+
},
|
| 124 |
+
{
|
| 125 |
+
"model_name": "gemini-2.0-flash",
|
| 126 |
+
"url": "https://ai.google.dev/gemini-api/docs/models#gemini-2.0-flash",
|
| 127 |
+
"organizer": "Google",
|
| 128 |
+
"license": "Proprietary",
|
| 129 |
+
"MLE-Lite_Elo": 847,
|
| 130 |
+
"Tabular_Elo": 923,
|
| 131 |
+
"NLP_Elo": 860,
|
| 132 |
+
"CV_Elo": 978,
|
| 133 |
+
"Overall": 895,
|
| 134 |
+
},
|
| 135 |
+
{
|
| 136 |
+
"model_name": "gemini-2.0-pro",
|
| 137 |
+
"url": "https://blog.google/technology/google-deepmind/gemini-model-updates-february-2025/",
|
| 138 |
+
"organizer": "Google",
|
| 139 |
+
"license": "Proprietary",
|
| 140 |
+
"MLE-Lite_Elo": 1064,
|
| 141 |
+
"Tabular_Elo": 1139,
|
| 142 |
+
"NLP_Elo": 1028,
|
| 143 |
+
"CV_Elo": 973,
|
| 144 |
+
"Overall": 1054,
|
| 145 |
+
},
|
| 146 |
+
{
|
| 147 |
+
"model_name": "gemini-2.5-pro",
|
| 148 |
+
"url": "https://deepmind.google/technologies/gemini/pro/",
|
| 149 |
+
"organizer": "Google",
|
| 150 |
+
"license": "Proprietary",
|
| 151 |
+
"MLE-Lite_Elo": 1257,
|
| 152 |
+
"Tabular_Elo": 1150,
|
| 153 |
+
"NLP_Elo": 1266,
|
| 154 |
+
"CV_Elo": 1177,
|
| 155 |
+
"Overall": 1214,
|
| 156 |
+
},
|
| 157 |
]
|
| 158 |
master_df = pd.DataFrame(data)
|
| 159 |
+
|
| 160 |
CATEGORIES = ["Overall", "MLE-Lite", "Tabular", "NLP", "CV"]
|
| 161 |
DEFAULT_CATEGORY = "Overall"
|
| 162 |
category_to_column = {
|
| 163 |
+
"MLE-Lite": "MLE-Lite_Elo",
|
| 164 |
+
"Tabular": "Tabular_Elo",
|
| 165 |
+
"NLP": "NLP_Elo",
|
| 166 |
+
"CV": "CV_Elo",
|
| 167 |
+
"Overall": "Overall",
|
| 168 |
}
|
| 169 |
|
| 170 |
+
# ---------------------------------------------------------------------------
|
| 171 |
+
# Helper to slice & rank the DataFrame when category radio changes
|
| 172 |
+
# ---------------------------------------------------------------------------
|
| 173 |
+
|
| 174 |
+
def update_leaderboard(category: str) -> pd.DataFrame:
|
| 175 |
+
"""Return a DataFrame limited to the selected category and sorted by score."""
|
| 176 |
+
score_column = category_to_column.get(category, category_to_column[DEFAULT_CATEGORY])
|
| 177 |
+
|
| 178 |
+
cols = ["model_name", "url", "organizer", "license", score_column]
|
| 179 |
+
df = master_df[cols].copy()
|
| 180 |
+
df = df.sort_values(score_column, ascending=False).reset_index(drop=True)
|
| 181 |
+
df.insert(0, "Rank", df.index + 1)
|
| 182 |
+
|
| 183 |
+
# Convert model name → clickable link (HTML will be rendered in the table)
|
| 184 |
+
df["Model"] = df.apply(
|
| 185 |
+
lambda r: (
|
| 186 |
+
f"<a href='{r.url if pd.notna(r.url) else '#'}' target='_blank' "
|
| 187 |
+
f"class='model-link'>{r.model_name}</a>"
|
| 188 |
+
),
|
| 189 |
+
axis=1,
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 190 |
)
|
| 191 |
|
| 192 |
+
df = df.rename(columns={
|
| 193 |
+
score_column: "Elo Score",
|
| 194 |
+
"organizer": "Organizer",
|
| 195 |
+
"license": "License",
|
| 196 |
+
})
|
| 197 |
+
return df[["Rank", "Model", "Organizer", "License", "Elo Score"]]
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 198 |
|
| 199 |
+
# ---------------------------------------------------------------------------
|
| 200 |
+
# Basic placeholder DataFrames for the (currently disabled) evaluation queue
|
| 201 |
+
# ---------------------------------------------------------------------------
|
| 202 |
+
print("Warning: evaluation queue fetching is disabled/mocked.")
|
| 203 |
+
empty_queue_df = pd.DataFrame(columns=["Model", "Status", "Requested", "Started"])
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 204 |
|
| 205 |
+
# ---------------------------------------------------------------------------
|
| 206 |
+
# Helper for HF Spaces restart (optional)
|
| 207 |
+
# ---------------------------------------------------------------------------
|
| 208 |
|
| 209 |
+
def restart_space():
|
| 210 |
+
"""Restart the current Hugging Face Space (if running in one)."""
|
| 211 |
+
print(f"Attempting to restart space: {REPO_ID}")
|
| 212 |
+
# Insert actual restart logic if needed.
|
| 213 |
+
|
| 214 |
+
# ---------------------------------------------------------------------------
|
| 215 |
+
# CSS: project CSS (custom_css) + enhanced overrides
|
| 216 |
+
# ---------------------------------------------------------------------------
|
| 217 |
+
|
| 218 |
+
# --- CLEAN introduction typography override (this replaces the buggy version) ---
|
| 219 |
+
intro_css = """
|
| 220 |
+
/* --------------------------------------------------
|
| 221 |
+
INTRODUCTION BLOCK (font, size, alignment)
|
| 222 |
+
-------------------------------------------------- */
|
| 223 |
+
.introduction-wrapper, .introduction-section {
|
| 224 |
+
font-family: Georgia, serif;
|
| 225 |
+
font-size: 1.4rem !important; /* ≈22–23 px */
|
| 226 |
+
line-height: 1.75;
|
| 227 |
+
color: #344054;
|
| 228 |
+
text-align: center;
|
| 229 |
+
max-width: 900px;
|
| 230 |
+
margin: 0 auto 3rem auto;
|
| 231 |
}
|
| 232 |
+
.introduction-wrapper p, .introduction-section p {
|
|
|
|
| 233 |
margin-bottom: 1rem;
|
|
|
|
| 234 |
}
|
| 235 |
+
@media (max-width: 768px) {
|
| 236 |
+
.introduction-wrapper, .introduction-section {
|
| 237 |
+
font-size: 1.2rem !important;
|
| 238 |
+
}
|
| 239 |
}
|
| 240 |
+
"""
|
| 241 |
+
|
| 242 |
+
# --- Existing CSS (base layout, table, etc.) ---
|
| 243 |
+
base_css = """
|
| 244 |
+
/* Base & layout overrides (truncated for brevity) */
|
| 245 |
+
body {
|
| 246 |
+
font-family: -apple-system, BlinkMacSystemFont, 'Segoe UI', Roboto, Oxygen,
|
| 247 |
+
Ubuntu, Cantarell, 'Open Sans', 'Helvetica Neue', sans-serif;
|
| 248 |
font-size: 1.3em;
|
| 249 |
+
line-height: 1.6;
|
| 250 |
+
background-color: #f8f9fa;
|
| 251 |
+
color: #343a40;
|
| 252 |
}
|
| 253 |
|
| 254 |
+
.gradio-container {
|
| 255 |
+
max-width: 1200px !important;
|
| 256 |
+
margin: 0 auto !important;
|
| 257 |
+
padding: 2rem !important;
|
|
|
|
| 258 |
}
|
| 259 |
|
| 260 |
+
#main-leaderboard-title {
|
| 261 |
+
font-size: 3.2em;
|
| 262 |
+
font-weight: 700;
|
| 263 |
+
color: #212529;
|
| 264 |
+
text-align: center;
|
| 265 |
+
margin-bottom: 1.5rem;
|
| 266 |
+
padding-bottom: 0.5rem;
|
| 267 |
+
border-bottom: 2px solid #dee2e6;
|
| 268 |
}
|
| 269 |
+
|
| 270 |
+
/* Leaderboard table (only key parts kept) */
|
| 271 |
#leaderboard-table th {
|
| 272 |
+
background-color: #e9ecef;
|
| 273 |
font-size: 1.3em;
|
| 274 |
+
font-weight: 500;
|
|
|
|
|
|
|
|
|
|
|
|
|
| 275 |
}
|
| 276 |
#leaderboard-table td {
|
| 277 |
font-size: 1.1em;
|
|
|
|
|
|
|
|
|
|
| 278 |
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 279 |
#leaderboard-table .model-link {
|
| 280 |
+
color: #0056b3;
|
|
|
|
| 281 |
font-weight: 500;
|
| 282 |
text-decoration: none;
|
| 283 |
}
|
| 284 |
#leaderboard-table .model-link:hover {
|
|
|
|
| 285 |
text-decoration: underline;
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 286 |
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 287 |
"""
|
| 288 |
|
| 289 |
+
# Concatenate user‑defined, base and intro CSS (order = priority)
|
| 290 |
+
final_css = f"{custom_css}\n{base_css}\n{intro_css}"
|
| 291 |
+
|
| 292 |
+
# ---------------------------------------------------------------------------
|
| 293 |
+
# Build the Gradio UI
|
| 294 |
+
# ---------------------------------------------------------------------------
|
| 295 |
|
| 296 |
+
demo = gr.Blocks(css=final_css, theme=gr.themes.Soft())
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 297 |
|
| 298 |
with demo:
|
| 299 |
+
# Title
|
| 300 |
gr.HTML(TITLE)
|
| 301 |
|
| 302 |
+
# Introduction (Markdown wrapped so CSS can target .introduction-wrapper)
|
|
|
|
| 303 |
with gr.Row():
|
| 304 |
+
gr.Markdown(INTRODUCTION_TEXT, elem_classes="introduction-wrapper")
|
| 305 |
|
| 306 |
+
with gr.Tabs(elem_classes="tab-buttons"):
|
| 307 |
+
# ------------------ Leaderboard tab ------------------
|
| 308 |
+
with gr.TabItem("🏅 MLE-Dojo Benchmark", id=0):
|
| 309 |
with gr.Column():
|
|
|
|
| 310 |
gr.Markdown("## Model Elo Rankings by Category", elem_classes="markdown-text")
|
| 311 |
+
|
| 312 |
category_selector = gr.Radio(
|
| 313 |
choices=CATEGORIES,
|
| 314 |
+
label="Select Category:",
|
| 315 |
value=DEFAULT_CATEGORY,
|
| 316 |
interactive=True,
|
| 317 |
+
elem_classes="gradio-radio",
|
| 318 |
)
|
| 319 |
+
|
| 320 |
leaderboard_df_component = gr.Dataframe(
|
| 321 |
value=update_leaderboard(DEFAULT_CATEGORY),
|
| 322 |
+
headers=["Rank", "Model", "Organizer", "License", "Elo Score"],
|
| 323 |
datatype=["number", "html", "str", "str", "number"],
|
| 324 |
interactive=False,
|
| 325 |
row_count=(len(master_df), "fixed"),
|
| 326 |
col_count=(5, "fixed"),
|
| 327 |
wrap=True,
|
| 328 |
+
elem_id="leaderboard-table",
|
| 329 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 330 |
|
| 331 |
+
category_selector.change(update_leaderboard, category_selector, leaderboard_df_component)
|
|
|
|
|
|
|
|
|
|
| 332 |
|
| 333 |
+
# ------------------ About tab ------------------
|
| 334 |
+
with gr.TabItem("📝 About", id=1):
|
| 335 |
+
gr.Markdown(LLM_BENCHMARKS_TEXT, elem_classes="markdown-text")
|
| 336 |
|
| 337 |
+
# Citation accordion (bottom of page)
|
| 338 |
+
with gr.Accordion("📙 Citation", open=False, elem_classes="gradio-accordion"):
|
| 339 |
+
gr.Textbox(
|
| 340 |
value=CITATION_BUTTON_TEXT,
|
| 341 |
label=CITATION_BUTTON_LABEL,
|
| 342 |
+
lines=8,
|
| 343 |
+
elem_id="citation-button",
|
| 344 |
show_copy_button=True,
|
| 345 |
)
|
| 346 |
|
| 347 |
+
# ---------------------------------------------------------------------------
|
| 348 |
+
# Scheduler (optional) & launch
|
| 349 |
+
# ---------------------------------------------------------------------------
|
| 350 |
if __name__ == "__main__":
|
| 351 |
try:
|
| 352 |
+
if callable(restart_space) and REPO_ID != "your/space-id":
|
| 353 |
+
scheduler = BackgroundScheduler()
|
| 354 |
+
scheduler.add_job(restart_space, "interval", seconds=1800)
|
| 355 |
+
scheduler.start()
|
| 356 |
+
print("Scheduler started for space restart.")
|
|
|
|
|
|
|
|
|
|
| 357 |
else:
|
| 358 |
+
print("Space restart scheduler not started (no REPO_ID or restart function).")
|
| 359 |
+
except Exception as exc:
|
| 360 |
+
print(f"Scheduler init failed: {exc}")
|
| 361 |
|
| 362 |
+
print("Launching Gradio app…")
|
| 363 |
+
demo.launch()
|
|
|