Spaces:
Running
Running
Update app.py
Browse files
app.py
CHANGED
@@ -3,55 +3,49 @@ import pandas as pd
|
|
3 |
from apscheduler.schedulers.background import BackgroundScheduler
|
4 |
|
5 |
"""
|
6 |
-
MLE‑Dojo Benchmark Leaderboard —
|
7 |
-
|
8 |
-
|
9 |
-
|
10 |
-
|
11 |
-
|
12 |
-
|
13 |
-
|
14 |
-
|
15 |
-
|
16 |
-
3. Tailor any of the placeholder values (TITLE, INTRODUCTION_TEXT, etc.) to your
|
17 |
-
project or import them from your own `src` package — the try/except block at
|
18 |
-
the top handles either workflow gracefully.
|
19 |
"""
|
20 |
|
21 |
# ---------------------------------------------------------------------------
|
22 |
-
#
|
23 |
# ---------------------------------------------------------------------------
|
24 |
try:
|
25 |
from src.about import (
|
26 |
CITATION_BUTTON_LABEL,
|
27 |
CITATION_BUTTON_TEXT,
|
28 |
-
EVALUATION_QUEUE_TEXT,
|
29 |
INTRODUCTION_TEXT,
|
30 |
LLM_BENCHMARKS_TEXT,
|
31 |
TITLE,
|
32 |
)
|
33 |
-
from src.display.css_html_js import custom_css # optional
|
34 |
from src.envs import REPO_ID
|
35 |
from src.submission.submit import add_new_eval
|
36 |
-
print("✅ Imported UI copy & helpers from src package.")
|
37 |
except ImportError:
|
38 |
-
|
39 |
CITATION_BUTTON_LABEL = "Citation"
|
40 |
CITATION_BUTTON_TEXT = "Please cite us if you use this benchmark…"
|
41 |
-
|
42 |
-
|
43 |
-
LLM_BENCHMARKS_TEXT = "Further details about tasks, metrics, and evaluation pipelines."
|
44 |
TITLE = (
|
45 |
-
"<h1 class='hero-title gradient-text'
|
46 |
-
"<p class='subtitle'>Interactive, reproducible & community‑driven ML
|
47 |
)
|
48 |
-
custom_css = ""
|
49 |
REPO_ID = "your/space-id"
|
50 |
def add_new_eval(*_):
|
51 |
return "Submission placeholder."
|
52 |
|
53 |
# ---------------------------------------------------------------------------
|
54 |
-
# Data
|
55 |
# ---------------------------------------------------------------------------
|
56 |
data = [
|
57 |
{"model_name": "gpt-4o-mini", "url": "https://openai.com/index/gpt-4o-mini-advancing-cost-efficient-intelligence/", "organizer": "OpenAI", "license": "Proprietary", "MLE-Lite_Elo": 753, "Tabular_Elo": 839, "NLP_Elo": 758, "CV_Elo": 754, "Overall": 778},
|
@@ -66,7 +60,7 @@ data = [
|
|
66 |
master_df = pd.DataFrame(data)
|
67 |
|
68 |
# ---------------------------------------------------------------------------
|
69 |
-
#
|
70 |
# ---------------------------------------------------------------------------
|
71 |
CATEGORIES = ["Overall", "MLE-Lite", "Tabular", "NLP", "CV"]
|
72 |
DEFAULT_CATEGORY = "Overall"
|
@@ -78,130 +72,75 @@ CATEGORY_MAP = {
|
|
78 |
"CV": "CV_Elo",
|
79 |
}
|
80 |
|
81 |
-
# ---------------------------------------------------------------------------
|
82 |
-
# Leaderboard Update Routine
|
83 |
-
# ---------------------------------------------------------------------------
|
84 |
-
|
85 |
def update_leaderboard(category: str, ascending: bool):
|
86 |
-
|
87 |
-
score_col = CATEGORY_MAP.get(category, CATEGORY_MAP[DEFAULT_CATEGORY])
|
88 |
-
|
89 |
df = (
|
90 |
-
master_df[["model_name", "url", "organizer", "license",
|
91 |
-
.sort_values(by=
|
92 |
.reset_index(drop=True)
|
93 |
)
|
94 |
-
|
95 |
-
# Add Rank & hyperlink the model name
|
96 |
df.insert(0, "Rank", df.index + 1)
|
97 |
-
df["Model"] = (
|
98 |
-
|
99 |
-
)
|
100 |
-
|
101 |
-
df.rename(columns={
|
102 |
-
"organizer": "Organizer",
|
103 |
-
"license": "License",
|
104 |
-
score_col: "Elo Score",
|
105 |
-
}, inplace=True)
|
106 |
-
|
107 |
return df[["Rank", "Model", "Organizer", "License", "Elo Score"]]
|
108 |
|
109 |
# ---------------------------------------------------------------------------
|
110 |
-
#
|
111 |
# ---------------------------------------------------------------------------
|
112 |
custom_css += """
|
|
|
113 |
@import url('https://fonts.googleapis.com/css2?family=Inter:wght@400;500;600;700;800&display=swap');
|
114 |
-
|
115 |
-
|
116 |
-
font-
|
117 |
-
font-size: 17px !important; /* slightly larger default */
|
118 |
-
color: #1f2937;
|
119 |
-
background-color: #f9fafb;
|
120 |
line-height: 1.55;
|
|
|
|
|
121 |
}
|
122 |
|
123 |
-
/*
|
124 |
.gradient-text {
|
125 |
-
background:
|
126 |
-
-webkit-background-clip: text;
|
127 |
-
-webkit-text-fill-color: transparent;
|
128 |
}
|
129 |
|
130 |
-
/* Markdown
|
131 |
-
.markdown-text h2
|
132 |
-
font-weight: 600;
|
133 |
-
margin-top: 1.2em;
|
134 |
-
}
|
135 |
|
136 |
-
/* Radio
|
137 |
-
.gr-radio, .gr-checkbox
|
138 |
-
|
139 |
-
border-radius: 0.5rem;
|
140 |
-
background-color: #ffffff;
|
141 |
-
box-shadow: 0 1px 2px rgba(0,0,0,0.06);
|
142 |
-
}
|
143 |
|
144 |
-
/*
|
145 |
-
#leaderboard-table table
|
146 |
-
|
147 |
-
|
148 |
-
}
|
149 |
-
#leaderboard-table
|
150 |
-
background-color: #e2e8f0;
|
151 |
-
font-weight: 600;
|
152 |
-
text-transform: uppercase;
|
153 |
-
font-size: 0.85rem;
|
154 |
-
letter-spacing: 0.03em;
|
155 |
-
padding: 0.6em;
|
156 |
-
}
|
157 |
-
#leaderboard-table td {
|
158 |
-
padding: 0.55em 0.6em;
|
159 |
-
vertical-align: top;
|
160 |
-
}
|
161 |
-
#leaderboard-table tr:nth-child(even) { background-color: #f8fafc; }
|
162 |
-
#leaderboard-table tr:hover { background-color: #eef2ff; }
|
163 |
|
164 |
/* Links */
|
165 |
-
a
|
166 |
-
a:hover { text-decoration: underline; }
|
167 |
|
168 |
-
/* Accordion
|
169 |
-
.gr-accordion .label
|
170 |
-
font-weight: 600;
|
171 |
-
font-size: 1rem;
|
172 |
-
}
|
173 |
"""
|
174 |
|
175 |
# ---------------------------------------------------------------------------
|
176 |
-
# Gradio App
|
177 |
# ---------------------------------------------------------------------------
|
|
|
178 |
|
179 |
-
|
180 |
-
primary_hue="indigo",
|
181 |
-
neutral_hue="slate",
|
182 |
-
font=["Inter", "Helvetica Neue", "Arial", "sans-serif"],
|
183 |
-
))
|
184 |
-
|
185 |
-
with demo:
|
186 |
gr.HTML(TITLE)
|
187 |
gr.Markdown(INTRODUCTION_TEXT, elem_classes="markdown-text")
|
188 |
|
189 |
with gr.Tabs():
|
190 |
-
|
191 |
-
|
192 |
-
gr.Markdown("### Model Elo Rankings by Category")
|
193 |
with gr.Row():
|
194 |
-
|
195 |
-
|
196 |
-
|
197 |
-
label="Category",
|
198 |
-
interactive=True,
|
199 |
-
)
|
200 |
-
order_checkbox = gr.Checkbox(
|
201 |
-
label="⬆️ Ascending order (lower Elo first)",
|
202 |
-
value=False,
|
203 |
-
)
|
204 |
-
leaderboard_table = gr.Dataframe(
|
205 |
value=update_leaderboard(DEFAULT_CATEGORY, False),
|
206 |
headers=["Rank", "Model", "Organizer", "License", "Elo Score"],
|
207 |
datatype=["number", "html", "str", "str", "number"],
|
@@ -210,42 +149,26 @@ with demo:
|
|
210 |
interactive=False,
|
211 |
elem_id="leaderboard-table",
|
212 |
)
|
213 |
-
|
214 |
-
|
215 |
-
order_checkbox.change(update_leaderboard, [category_selector, order_checkbox], leaderboard_table)
|
216 |
|
217 |
-
# ---------- About Tab ----------
|
218 |
with gr.TabItem("ℹ️ About"):
|
219 |
gr.Markdown(LLM_BENCHMARKS_TEXT, elem_classes="markdown-text")
|
220 |
|
221 |
-
# ---------- (Optional) Submit Tab ----------
|
222 |
-
# You can re‑enable this section when your `add_new_eval()` & REPO_ID are ready.
|
223 |
-
# with gr.TabItem("🚀 Submit"):
|
224 |
-
# pass
|
225 |
-
|
226 |
-
# ---------- Citation Accordion ----------
|
227 |
with gr.Accordion("📖 Citation", open=False):
|
228 |
-
gr.Textbox(
|
229 |
-
value=CITATION_BUTTON_TEXT,
|
230 |
-
label=CITATION_BUTTON_LABEL,
|
231 |
-
lines=10,
|
232 |
-
show_copy_button=True,
|
233 |
-
)
|
234 |
|
235 |
# ---------------------------------------------------------------------------
|
236 |
-
#
|
237 |
# ---------------------------------------------------------------------------
|
238 |
|
239 |
def restart_space():
|
240 |
-
print(f"
|
241 |
-
# Example: `HfApi().restart_space(repo_id=REPO_ID)`
|
242 |
|
243 |
if __name__ == "__main__":
|
244 |
if REPO_ID != "your/space-id":
|
245 |
scheduler = BackgroundScheduler()
|
246 |
scheduler.add_job(restart_space, "interval", seconds=1800)
|
247 |
scheduler.start()
|
248 |
-
|
249 |
-
|
250 |
-
print("🚀 Launching Gradio app…")
|
251 |
-
demo.launch()
|
|
|
3 |
from apscheduler.schedulers.background import BackgroundScheduler
|
4 |
|
5 |
"""
|
6 |
+
MLE‑Dojo Benchmark Leaderboard — Dark Elegance Edition
|
7 |
+
=====================================================
|
8 |
+
A refined, low‑glare UI with larger table text, richer surface layering, and a
|
9 |
+
subtle neon accent that pops on dark slate backgrounds.
|
10 |
+
|
11 |
+
HOW TO
|
12 |
+
------
|
13 |
+
* `pip install gradio pandas apscheduler`
|
14 |
+
* `python mle_dojo_leaderboard_app.py`
|
15 |
+
* Replace placeholder copy (TITLE …) with your own or keep the try/except.
|
|
|
|
|
|
|
16 |
"""
|
17 |
|
18 |
# ---------------------------------------------------------------------------
|
19 |
+
# Import app copy (falls back to placeholders if src/ is absent)
|
20 |
# ---------------------------------------------------------------------------
|
21 |
try:
|
22 |
from src.about import (
|
23 |
CITATION_BUTTON_LABEL,
|
24 |
CITATION_BUTTON_TEXT,
|
|
|
25 |
INTRODUCTION_TEXT,
|
26 |
LLM_BENCHMARKS_TEXT,
|
27 |
TITLE,
|
28 |
)
|
29 |
+
from src.display.css_html_js import custom_css # optional extra rules
|
30 |
from src.envs import REPO_ID
|
31 |
from src.submission.submit import add_new_eval
|
|
|
32 |
except ImportError:
|
33 |
+
# ── Placeholders ───────────────────────────────────────────────────────────
|
34 |
CITATION_BUTTON_LABEL = "Citation"
|
35 |
CITATION_BUTTON_TEXT = "Please cite us if you use this benchmark…"
|
36 |
+
INTRODUCTION_TEXT = "Welcome to the **MLE‑Dojo Benchmark Leaderboard** — compare LLM agents across realistic ML engineering tasks."
|
37 |
+
LLM_BENCHMARKS_TEXT = "Further details about tasks, metrics and evaluation pipelines."
|
|
|
38 |
TITLE = (
|
39 |
+
"<h1 class='hero-title gradient-text'>🏆 MLE‑Dojo Benchmark Leaderboard</h1>"
|
40 |
+
"<p class='subtitle'>Interactive, reproducible & community‑driven ML‑agent benchmarking</p>"
|
41 |
)
|
42 |
+
custom_css = ""
|
43 |
REPO_ID = "your/space-id"
|
44 |
def add_new_eval(*_):
|
45 |
return "Submission placeholder."
|
46 |
|
47 |
# ---------------------------------------------------------------------------
|
48 |
+
# Data
|
49 |
# ---------------------------------------------------------------------------
|
50 |
data = [
|
51 |
{"model_name": "gpt-4o-mini", "url": "https://openai.com/index/gpt-4o-mini-advancing-cost-efficient-intelligence/", "organizer": "OpenAI", "license": "Proprietary", "MLE-Lite_Elo": 753, "Tabular_Elo": 839, "NLP_Elo": 758, "CV_Elo": 754, "Overall": 778},
|
|
|
60 |
master_df = pd.DataFrame(data)
|
61 |
|
62 |
# ---------------------------------------------------------------------------
|
63 |
+
# Helpers
|
64 |
# ---------------------------------------------------------------------------
|
65 |
CATEGORIES = ["Overall", "MLE-Lite", "Tabular", "NLP", "CV"]
|
66 |
DEFAULT_CATEGORY = "Overall"
|
|
|
72 |
"CV": "CV_Elo",
|
73 |
}
|
74 |
|
|
|
|
|
|
|
|
|
75 |
def update_leaderboard(category: str, ascending: bool):
|
76 |
+
col = CATEGORY_MAP.get(category, CATEGORY_MAP[DEFAULT_CATEGORY])
|
|
|
|
|
77 |
df = (
|
78 |
+
master_df[["model_name", "url", "organizer", "license", col]]
|
79 |
+
.sort_values(by=col, ascending=ascending)
|
80 |
.reset_index(drop=True)
|
81 |
)
|
|
|
|
|
82 |
df.insert(0, "Rank", df.index + 1)
|
83 |
+
df["Model"] = df.apply(lambda r: f"<a href='{r.url}' target='_blank'>{r.model_name}</a>", axis=1)
|
84 |
+
df.rename(columns={"organizer": "Organizer", "license": "License", col: "Elo Score"}, inplace=True)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
85 |
return df[["Rank", "Model", "Organizer", "License", "Elo Score"]]
|
86 |
|
87 |
# ---------------------------------------------------------------------------
|
88 |
+
# Dark‑mode CSS & Larger Table Fonts
|
89 |
# ---------------------------------------------------------------------------
|
90 |
custom_css += """
|
91 |
+
/* ————— Core Typography ————— */
|
92 |
@import url('https://fonts.googleapis.com/css2?family=Inter:wght@400;500;600;700;800&display=swap');
|
93 |
+
html,body {
|
94 |
+
font-family: 'Inter', sans-serif !important;
|
95 |
+
font-size: 18px !important;
|
|
|
|
|
|
|
96 |
line-height: 1.55;
|
97 |
+
color: #e2e8f0;
|
98 |
+
background:#0f172a;
|
99 |
}
|
100 |
|
101 |
+
/* Accent gradient for titles */
|
102 |
.gradient-text {
|
103 |
+
background:linear-gradient(90deg,#6366f1 0%,#06b6d4 100%);
|
104 |
+
-webkit-background-clip:text; -webkit-text-fill-color:transparent;
|
|
|
105 |
}
|
106 |
|
107 |
+
/* Markdown headings */
|
108 |
+
.markdown-text h2{font-weight:600;margin-top:1.3em;color:#f1f5f9;}
|
|
|
|
|
|
|
109 |
|
110 |
+
/* Radio & checkbox containers */
|
111 |
+
.gr-radio, .gr-checkbox{background:#1e293b;border-radius:8px;padding:6px 12px;box-shadow:0 1px 3px rgba(0,0,0,.4);}
|
112 |
+
.gr-radio input:checked+label, .gr-checkbox input:checked+label{color:#38bdf8;}
|
|
|
|
|
|
|
|
|
113 |
|
114 |
+
/* Table Styling */
|
115 |
+
#leaderboard-table table{width:100%;border-collapse:collapse;background:#1e293b;border-radius:8px;overflow:hidden;}
|
116 |
+
#leaderboard-table th{background:#334155;font-size:0.9rem;font-weight:600;padding:0.7em;color:#f1f5f9;text-transform:uppercase;letter-spacing:.04em;}
|
117 |
+
#leaderboard-table td{padding:0.6em;font-size:1.05rem;border-top:1px solid #334155;}
|
118 |
+
#leaderboard-table tr:nth-child(even){background:#1c2431;}
|
119 |
+
#leaderboard-table tr:hover{background:#475569;}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
120 |
|
121 |
/* Links */
|
122 |
+
a{color:#38bdf8;} a:hover{text-decoration:underline;}
|
|
|
123 |
|
124 |
+
/* Accordion */
|
125 |
+
.gr-accordion .label{font-weight:600;font-size:1rem;color:#f1f5f9;}
|
|
|
|
|
|
|
126 |
"""
|
127 |
|
128 |
# ---------------------------------------------------------------------------
|
129 |
+
# Gradio App
|
130 |
# ---------------------------------------------------------------------------
|
131 |
+
app = gr.Blocks(css=custom_css, theme=gr.themes.Soft(primary_hue="sky", neutral_hue="slate", font=["Inter",]))
|
132 |
|
133 |
+
with app:
|
|
|
|
|
|
|
|
|
|
|
|
|
134 |
gr.HTML(TITLE)
|
135 |
gr.Markdown(INTRODUCTION_TEXT, elem_classes="markdown-text")
|
136 |
|
137 |
with gr.Tabs():
|
138 |
+
with gr.TabItem("🏅 Leaderboard"):
|
139 |
+
gr.Markdown("### Model Elo Rankings by Category", elem_classes="markdown-text")
|
|
|
140 |
with gr.Row():
|
141 |
+
category_radio = gr.Radio(CATEGORIES, value=DEFAULT_CATEGORY, label="Category")
|
142 |
+
asc_check = gr.Checkbox(label="⬆️ Ascending order", value=False)
|
143 |
+
board = gr.Dataframe(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
144 |
value=update_leaderboard(DEFAULT_CATEGORY, False),
|
145 |
headers=["Rank", "Model", "Organizer", "License", "Elo Score"],
|
146 |
datatype=["number", "html", "str", "str", "number"],
|
|
|
149 |
interactive=False,
|
150 |
elem_id="leaderboard-table",
|
151 |
)
|
152 |
+
category_radio.change(update_leaderboard, [category_radio, asc_check], board)
|
153 |
+
asc_check.change(update_leaderboard, [category_radio, asc_check], board)
|
|
|
154 |
|
|
|
155 |
with gr.TabItem("ℹ️ About"):
|
156 |
gr.Markdown(LLM_BENCHMARKS_TEXT, elem_classes="markdown-text")
|
157 |
|
|
|
|
|
|
|
|
|
|
|
|
|
158 |
with gr.Accordion("📖 Citation", open=False):
|
159 |
+
gr.Textbox(value=CITATION_BUTTON_TEXT, label=CITATION_BUTTON_LABEL, lines=10, show_copy_button=True)
|
|
|
|
|
|
|
|
|
|
|
160 |
|
161 |
# ---------------------------------------------------------------------------
|
162 |
+
# Optional scheduler (HF Space restarts)
|
163 |
# ---------------------------------------------------------------------------
|
164 |
|
165 |
def restart_space():
|
166 |
+
print(f"Restarting space → {REPO_ID}")
|
|
|
167 |
|
168 |
if __name__ == "__main__":
|
169 |
if REPO_ID != "your/space-id":
|
170 |
scheduler = BackgroundScheduler()
|
171 |
scheduler.add_job(restart_space, "interval", seconds=1800)
|
172 |
scheduler.start()
|
173 |
+
print("Launching app…")
|
174 |
+
app.launch()
|
|
|
|