Spaces:
Running
Running
Update app.py
Browse files
app.py
CHANGED
@@ -65,12 +65,13 @@ def update_leaderboard(category):
|
|
65 |
|
66 |
# --- Dark Mode Styling ---
|
67 |
font_size_css = """
|
68 |
-
body { font-size:
|
69 |
"""
|
70 |
custom_css += font_size_css
|
71 |
|
72 |
dark_css = """
|
73 |
@import url('https://fonts.googleapis.com/css2?family=Inter:wght@400;500;600;700&display=swap');
|
|
|
74 |
|
75 |
body {
|
76 |
font-family: 'Inter', sans-serif;
|
@@ -90,213 +91,96 @@ body {
|
|
90 |
}
|
91 |
.hero-section h1 {
|
92 |
margin: 0;
|
93 |
-
font-size: 2.
|
94 |
font-weight: 700 !important;
|
95 |
-
display: flex;
|
96 |
-
align-items: center;
|
97 |
-
justify-content: center;
|
98 |
-
}
|
99 |
-
.hero-section h1 i {
|
100 |
-
margin-right: 0.5rem;
|
101 |
-
color: gold;
|
102 |
}
|
103 |
.hero-section h2 {
|
104 |
margin: .5rem 0 0 !important;
|
105 |
-
font-size: 1.
|
106 |
font-weight: 400 !important;
|
107 |
opacity: 0.8;
|
108 |
}
|
109 |
|
110 |
/* Tab Buttons */
|
111 |
-
.tab-
|
112 |
border-radius: 20px !important;
|
113 |
-
padding: 0.
|
114 |
-
margin-right: 0.
|
115 |
background: #1e1e1e !important;
|
116 |
color: #e0e0e0 !important;
|
117 |
border: none !important;
|
118 |
-
transition:
|
119 |
-
font-weight:
|
120 |
-
font-size: 1.2rem !important;
|
121 |
-
letter-spacing: 0.02em !important;
|
122 |
-
box-shadow: 0 2px 4px rgba(0,0,0,0.2) !important;
|
123 |
}
|
124 |
-
.tab-
|
125 |
background: #2c2c2c !important;
|
126 |
-
transform: translateY(-2px) !important;
|
127 |
-
box-shadow: 0 4px 8px rgba(0,0,0,0.3) !important;
|
128 |
}
|
129 |
-
.tab-
|
130 |
-
background:
|
131 |
color: #fff !important;
|
132 |
-
box-shadow: 0 2px 6px rgba(0,0,0,0.4) !important;
|
133 |
}
|
134 |
|
135 |
-
/* Category Selector */
|
136 |
-
#category-selector .wrap {
|
137 |
-
display: flex;
|
138 |
-
flex-direction: column;
|
139 |
-
gap: 0.75rem;
|
140 |
-
}
|
141 |
-
#category-selector .label-wrap {
|
142 |
-
font-size: 1.4rem !important;
|
143 |
-
font-weight: 600 !important;
|
144 |
-
color: #e0e0e0 !important;
|
145 |
-
letter-spacing: 0.02em !important;
|
146 |
-
margin-bottom: 0.5rem !important;
|
147 |
-
}
|
148 |
#category-selector input[type="radio"] { display: none; }
|
149 |
-
#category-selector .options {
|
150 |
-
display: flex;
|
151 |
-
flex-direction: row !important;
|
152 |
-
flex-wrap: wrap;
|
153 |
-
gap: 0.75rem;
|
154 |
-
justify-content: space-between;
|
155 |
-
}
|
156 |
#category-selector label {
|
157 |
display: inline-block;
|
158 |
-
padding: 0.
|
|
|
159 |
border-radius: 999px;
|
160 |
background: #1e1e1e;
|
161 |
cursor: pointer;
|
162 |
-
transition:
|
163 |
font-weight: 500;
|
164 |
color: #e0e0e0;
|
165 |
-
letter-spacing: 0.02em;
|
166 |
-
font-size: 1.15rem;
|
167 |
-
box-shadow: 0 2px 4px rgba(0,0,0,0.15);
|
168 |
-
text-align: center;
|
169 |
-
width: 18%;
|
170 |
-
margin: 0 !important;
|
171 |
-
}
|
172 |
-
#category-selector label:hover {
|
173 |
-
background: #2a2a2a;
|
174 |
-
transform: translateY(-2px);
|
175 |
-
box-shadow: 0 4px 8px rgba(0,0,0,0.2);
|
176 |
}
|
177 |
#category-selector input[type="radio"]:checked + label {
|
178 |
-
background:
|
179 |
color: #fff;
|
180 |
-
box-shadow: 0 2px 6px rgba(0,0,0,0.3);
|
181 |
-
font-weight: 600;
|
182 |
}
|
183 |
|
184 |
/* Table Styling */
|
185 |
-
.dataframe-container {
|
186 |
-
margin-top: 1.5rem;
|
187 |
-
border-radius: 0.75rem;
|
188 |
-
overflow: hidden;
|
189 |
-
box-shadow: 0 4px 12px rgba(0,0,0,0.3);
|
190 |
-
}
|
191 |
.dataframe-container table {
|
192 |
width: 100%;
|
193 |
border: none;
|
194 |
-
|
|
|
|
|
195 |
}
|
196 |
.dataframe-container table th {
|
197 |
-
background:
|
198 |
-
color: #
|
199 |
-
font-weight: 600;
|
200 |
-
font-size: 1.25rem;
|
201 |
-
text-align: left;
|
202 |
-
letter-spacing: 0.02em;
|
203 |
-
padding: 1.2rem 1.5rem;
|
204 |
-
position: sticky;
|
205 |
-
top: 0;
|
206 |
-
box-shadow: 0 2px 4px rgba(0,0,0,0.2);
|
207 |
-
z-index: 10;
|
208 |
}
|
209 |
.dataframe-container table tr:nth-child(odd) {
|
210 |
-
background-color: #
|
211 |
}
|
212 |
.dataframe-container table tr:nth-child(even) {
|
213 |
-
background-color: #
|
214 |
-
}
|
215 |
-
.dataframe-container table tr:hover {
|
216 |
-
background-color: #2c2c2c !important;
|
217 |
-
transition: background-color 0.2s ease;
|
218 |
}
|
219 |
-
.dataframe-container table td {
|
220 |
-
padding:
|
221 |
color: #e0e0e0;
|
222 |
-
font-size: 1.2rem;
|
223 |
-
letter-spacing: 0.01em;
|
224 |
-
border-bottom: 1px solid #333;
|
225 |
}
|
226 |
.dataframe-container table td a {
|
227 |
color: #8ab4f8;
|
228 |
text-decoration: none;
|
229 |
-
font-weight: 500;
|
230 |
-
transition: all 0.2s ease;
|
231 |
}
|
232 |
.dataframe-container table td a:hover {
|
233 |
color: #a3c9ff;
|
234 |
text-decoration: underline;
|
235 |
}
|
236 |
|
237 |
-
/* Column-specific styles */
|
238 |
-
.dataframe-container table td:first-child,
|
239 |
-
.dataframe-container table th:first-child {
|
240 |
-
text-align: center;
|
241 |
-
width: 80px;
|
242 |
-
}
|
243 |
-
.dataframe-container table td:last-child,
|
244 |
-
.dataframe-container table th:last-child {
|
245 |
-
text-align: center;
|
246 |
-
font-weight: 600;
|
247 |
-
font-size: 1.3rem;
|
248 |
-
color: #ffcc00;
|
249 |
-
}
|
250 |
-
|
251 |
/* Enable scrollbar */
|
252 |
#leaderboard-table .dataframe-container {
|
253 |
-
max-height:
|
254 |
overflow-y: auto !important;
|
255 |
}
|
256 |
-
|
257 |
-
/* Custom Scrollbar for WebKit browsers */
|
258 |
-
.dataframe-container::-webkit-scrollbar {
|
259 |
-
width: 10px;
|
260 |
-
}
|
261 |
-
.dataframe-container::-webkit-scrollbar-track {
|
262 |
-
background: #1a1a1a;
|
263 |
-
border-radius: 5px;
|
264 |
-
}
|
265 |
-
.dataframe-container::-webkit-scrollbar-thumb {
|
266 |
-
background: #444;
|
267 |
-
border-radius: 5px;
|
268 |
-
}
|
269 |
-
.dataframe-container::-webkit-scrollbar-thumb:hover {
|
270 |
-
background: #555;
|
271 |
-
}
|
272 |
-
|
273 |
-
/* Markdown text styling */
|
274 |
-
.markdown-text {
|
275 |
-
line-height: 1.6;
|
276 |
-
font-size: 1.2rem;
|
277 |
-
color: #d0d0d0;
|
278 |
-
}
|
279 |
-
.markdown-text h1, .markdown-text h2, .markdown-text h3 {
|
280 |
-
color: #fff;
|
281 |
-
font-weight: 600;
|
282 |
-
margin-top: 1.5rem;
|
283 |
-
margin-bottom: 1rem;
|
284 |
-
}
|
285 |
-
.markdown-text h2 {
|
286 |
-
font-size: 1.8rem !important;
|
287 |
-
border-bottom: 1px solid #333;
|
288 |
-
padding-bottom: 0.5rem;
|
289 |
-
}
|
290 |
-
.markdown-text p {
|
291 |
-
margin-bottom: 1rem;
|
292 |
-
}
|
293 |
"""
|
294 |
custom_css += dark_css
|
295 |
|
296 |
# --- Override Title ---
|
297 |
TITLE = """
|
298 |
-
<div class
|
299 |
-
<h1><i class
|
300 |
<h2>Improving LLM Agents for Machine Learning Engineering</h2>
|
301 |
</div>
|
302 |
"""
|
@@ -306,27 +190,24 @@ TITLE = """
|
|
306 |
demo = gr.Blocks(css=custom_css, theme=gr.themes.Base())
|
307 |
|
308 |
with demo:
|
309 |
-
# Inject FontAwesome CSS
|
310 |
gr.HTML("""
|
311 |
-
<
|
312 |
-
|
313 |
-
|
314 |
-
|
315 |
-
crossorigin
|
316 |
-
|
317 |
-
|
318 |
-
crossorigin="anonymous" referrerpolicy="no-referrer"></script>
|
319 |
-
</head>
|
320 |
-
""")
|
321 |
|
322 |
# Header & Intro
|
323 |
gr.HTML(TITLE)
|
324 |
gr.Markdown(INTRODUCTION_TEXT, elem_classes="markdown-text")
|
325 |
|
326 |
-
# Tabs
|
327 |
-
with gr.Tabs(elem_classes="tab-
|
328 |
-
with gr.TabItem("
|
329 |
-
gr.Markdown("## Model Elo Rankings by Category"
|
330 |
category_selector = gr.Radio(
|
331 |
choices=CATEGORIES,
|
332 |
label="Select Category:",
|
@@ -350,7 +231,7 @@ with demo:
|
|
350 |
outputs=leaderboard_df
|
351 |
)
|
352 |
|
353 |
-
with gr.TabItem("
|
354 |
gr.Markdown(LLM_BENCHMARKS_TEXT, elem_classes="markdown-text")
|
355 |
|
356 |
# Citation Accordion
|
|
|
65 |
|
66 |
# --- Dark Mode Styling ---
|
67 |
font_size_css = """
|
68 |
+
body { font-size: 1em !important; }
|
69 |
"""
|
70 |
custom_css += font_size_css
|
71 |
|
72 |
dark_css = """
|
73 |
@import url('https://fonts.googleapis.com/css2?family=Inter:wght@400;500;600;700&display=swap');
|
74 |
+
@import url('https://cdnjs.cloudflare.com/ajax/libs/font-awesome/6.4.0/css/all.min.css');
|
75 |
|
76 |
body {
|
77 |
font-family: 'Inter', sans-serif;
|
|
|
91 |
}
|
92 |
.hero-section h1 {
|
93 |
margin: 0;
|
94 |
+
font-size: 2.5rem !important;
|
95 |
font-weight: 700 !important;
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
96 |
}
|
97 |
.hero-section h2 {
|
98 |
margin: .5rem 0 0 !important;
|
99 |
+
font-size: 1.25rem !important;
|
100 |
font-weight: 400 !important;
|
101 |
opacity: 0.8;
|
102 |
}
|
103 |
|
104 |
/* Tab Buttons */
|
105 |
+
.tab-buttons button {
|
106 |
border-radius: 20px !important;
|
107 |
+
padding: 0.5rem 1rem !important;
|
108 |
+
margin-right: 0.5rem !important;
|
109 |
background: #1e1e1e !important;
|
110 |
color: #e0e0e0 !important;
|
111 |
border: none !important;
|
112 |
+
transition: background 0.3s !important;
|
113 |
+
font-weight: 500 !important;
|
|
|
|
|
|
|
114 |
}
|
115 |
+
.tab-buttons button:hover {
|
116 |
background: #2c2c2c !important;
|
|
|
|
|
117 |
}
|
118 |
+
.tab-buttons button[aria-selected="true"] {
|
119 |
+
background: #444 !important;
|
120 |
color: #fff !important;
|
|
|
121 |
}
|
122 |
|
123 |
+
/* Category Selector Pills */
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
124 |
#category-selector input[type="radio"] { display: none; }
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
125 |
#category-selector label {
|
126 |
display: inline-block;
|
127 |
+
padding: 0.5rem 1rem;
|
128 |
+
margin-right: 0.5rem;
|
129 |
border-radius: 999px;
|
130 |
background: #1e1e1e;
|
131 |
cursor: pointer;
|
132 |
+
transition: background 0.3s, color 0.3s;
|
133 |
font-weight: 500;
|
134 |
color: #e0e0e0;
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
135 |
}
|
136 |
#category-selector input[type="radio"]:checked + label {
|
137 |
+
background: #444;
|
138 |
color: #fff;
|
|
|
|
|
139 |
}
|
140 |
|
141 |
/* Table Styling */
|
|
|
|
|
|
|
|
|
|
|
|
|
142 |
.dataframe-container table {
|
143 |
width: 100%;
|
144 |
border: none;
|
145 |
+
box-shadow: 0 2px 4px rgba(0,0,0,0.5);
|
146 |
+
border-radius: 0.5rem;
|
147 |
+
overflow: hidden;
|
148 |
}
|
149 |
.dataframe-container table th {
|
150 |
+
background: #2c2c2c;
|
151 |
+
color: #e0e0e0;
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
152 |
}
|
153 |
.dataframe-container table tr:nth-child(odd) {
|
154 |
+
background-color: #1e1e1e !important;
|
155 |
}
|
156 |
.dataframe-container table tr:nth-child(even) {
|
157 |
+
background-color: #252525 !important;
|
|
|
|
|
|
|
|
|
158 |
}
|
159 |
+
.dataframe-container table td, .dataframe-container table th {
|
160 |
+
padding: 0.75rem 1rem;
|
161 |
color: #e0e0e0;
|
|
|
|
|
|
|
162 |
}
|
163 |
.dataframe-container table td a {
|
164 |
color: #8ab4f8;
|
165 |
text-decoration: none;
|
|
|
|
|
166 |
}
|
167 |
.dataframe-container table td a:hover {
|
168 |
color: #a3c9ff;
|
169 |
text-decoration: underline;
|
170 |
}
|
171 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
172 |
/* Enable scrollbar */
|
173 |
#leaderboard-table .dataframe-container {
|
174 |
+
max-height: 400px !important;
|
175 |
overflow-y: auto !important;
|
176 |
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
177 |
"""
|
178 |
custom_css += dark_css
|
179 |
|
180 |
# --- Override Title ---
|
181 |
TITLE = """
|
182 |
+
<div class=\"hero-section\">
|
183 |
+
<h1><i class=\"fas fa-trophy\"></i> MLE-Dojo Benchmark Leaderboard</h1>
|
184 |
<h2>Improving LLM Agents for Machine Learning Engineering</h2>
|
185 |
</div>
|
186 |
"""
|
|
|
190 |
demo = gr.Blocks(css=custom_css, theme=gr.themes.Base())
|
191 |
|
192 |
with demo:
|
193 |
+
# Inject FontAwesome JS/CSS explicitly
|
194 |
gr.HTML("""
|
195 |
+
<link rel=\"stylesheet\"
|
196 |
+
href=\"https://cdnjs.cloudflare.com/ajax/libs/font-awesome/6.4.0/css/all.min.css\"
|
197 |
+
crossorigin=\"anonymous\" referrerpolicy=\"no-referrer\"/>
|
198 |
+
<script src=\"https://cdnjs.cloudflare.com/ajax/libs/font-awesome/6.4.0/js/all.min.js\"
|
199 |
+
crossorigin=\"anonymous\" referrerpolicy=\"no-referrer\"></script>
|
200 |
+
"""
|
201 |
+
)
|
|
|
|
|
|
|
202 |
|
203 |
# Header & Intro
|
204 |
gr.HTML(TITLE)
|
205 |
gr.Markdown(INTRODUCTION_TEXT, elem_classes="markdown-text")
|
206 |
|
207 |
+
# Tabs
|
208 |
+
with gr.Tabs(elem_classes="tab-buttons") as tabs:
|
209 |
+
with gr.TabItem("<i class='fas fa-list'></i> Leaderboard"):
|
210 |
+
gr.Markdown("## Model Elo Rankings by Category")
|
211 |
category_selector = gr.Radio(
|
212 |
choices=CATEGORIES,
|
213 |
label="Select Category:",
|
|
|
231 |
outputs=leaderboard_df
|
232 |
)
|
233 |
|
234 |
+
with gr.TabItem("<i class='fas fa-info-circle'></i> About"):
|
235 |
gr.Markdown(LLM_BENCHMARKS_TEXT, elem_classes="markdown-text")
|
236 |
|
237 |
# Citation Accordion
|