Spaces:
Running
Running
Update app.py
Browse files
app.py
CHANGED
@@ -3,20 +3,17 @@ import pandas as pd
|
|
3 |
from apscheduler.schedulers.background import BackgroundScheduler
|
4 |
|
5 |
"""
|
6 |
-
MLE‑Dojo Benchmark Leaderboard — Dark Elegance
|
7 |
-
|
8 |
-
|
9 |
-
|
10 |
-
|
11 |
-
|
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 (
|
20 |
# ---------------------------------------------------------------------------
|
21 |
try:
|
22 |
from src.about import (
|
@@ -26,15 +23,14 @@ try:
|
|
26 |
LLM_BENCHMARKS_TEXT,
|
27 |
TITLE,
|
28 |
)
|
29 |
-
from src.display.css_html_js import custom_css #
|
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
|
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>"
|
@@ -72,6 +68,8 @@ CATEGORY_MAP = {
|
|
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 = (
|
@@ -79,49 +77,45 @@ def update_leaderboard(category: str, ascending: bool):
|
|
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 |
-
#
|
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 |
-
|
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 |
-
/*
|
108 |
-
.
|
|
|
109 |
|
110 |
-
/* Radio & checkbox
|
111 |
-
.gr-radio
|
112 |
-
.gr-radio input:checked+label
|
113 |
|
114 |
-
/*
|
115 |
#leaderboard-table table{width:100%;border-collapse:collapse;background:#1e293b;border-radius:8px;overflow:hidden;}
|
116 |
-
#leaderboard-table th{background:#334155;font-size:
|
117 |
-
#leaderboard-table td{padding
|
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 |
|
@@ -136,18 +130,20 @@ with app:
|
|
136 |
|
137 |
with gr.Tabs():
|
138 |
with gr.TabItem("🏅 Leaderboard"):
|
139 |
-
gr.
|
140 |
with gr.Row():
|
141 |
category_radio = gr.Radio(CATEGORIES, value=DEFAULT_CATEGORY, label="Category")
|
142 |
-
asc_check = gr.Checkbox(label="⬆️
|
|
|
143 |
board = gr.Dataframe(
|
144 |
value=update_leaderboard(DEFAULT_CATEGORY, False),
|
145 |
headers=["Rank", "Model", "Organizer", "License", "Elo Score"],
|
146 |
-
datatype=["
|
147 |
row_count=(len(master_df), "fixed"),
|
148 |
col_count=(5, "fixed"),
|
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)
|
|
|
3 |
from apscheduler.schedulers.background import BackgroundScheduler
|
4 |
|
5 |
"""
|
6 |
+
MLE‑Dojo Benchmark Leaderboard — Dark Elegance v2
|
7 |
+
================================================
|
8 |
+
Upgrades:
|
9 |
+
* richer heading with emoji icon
|
10 |
+
* larger fonts throughout
|
11 |
+
* medal icons for top‑3 ranks
|
12 |
+
* scrollable leaderboard (fixed height)
|
|
|
|
|
|
|
13 |
"""
|
14 |
|
15 |
# ---------------------------------------------------------------------------
|
16 |
+
# Import app copy (fallback placeholders if src/ missing)
|
17 |
# ---------------------------------------------------------------------------
|
18 |
try:
|
19 |
from src.about import (
|
|
|
23 |
LLM_BENCHMARKS_TEXT,
|
24 |
TITLE,
|
25 |
)
|
26 |
+
from src.display.css_html_js import custom_css # may exist
|
27 |
from src.envs import REPO_ID
|
28 |
from src.submission.submit import add_new_eval
|
29 |
except ImportError:
|
|
|
30 |
CITATION_BUTTON_LABEL = "Citation"
|
31 |
CITATION_BUTTON_TEXT = "Please cite us if you use this benchmark…"
|
32 |
+
INTRODUCTION_TEXT = "Welcome to the **MLE‑Dojo Benchmark Leaderboard** — compare LLM agents across realistic ML‑engineering tasks."
|
33 |
+
LLM_BENCHMARKS_TEXT = "Further details about tasks, metrics, and evaluation pipelines."
|
34 |
TITLE = (
|
35 |
"<h1 class='hero-title gradient-text'>🏆 MLE‑Dojo Benchmark Leaderboard</h1>"
|
36 |
"<p class='subtitle'>Interactive, reproducible & community‑driven ML‑agent benchmarking</p>"
|
|
|
68 |
"CV": "CV_Elo",
|
69 |
}
|
70 |
|
71 |
+
MEDALS = {1: "🥇", 2: "🥈", 3: "🥉"}
|
72 |
+
|
73 |
def update_leaderboard(category: str, ascending: bool):
|
74 |
col = CATEGORY_MAP.get(category, CATEGORY_MAP[DEFAULT_CATEGORY])
|
75 |
df = (
|
|
|
77 |
.sort_values(by=col, ascending=ascending)
|
78 |
.reset_index(drop=True)
|
79 |
)
|
80 |
+
|
81 |
df.insert(0, "Rank", df.index + 1)
|
82 |
+
df["Rank"] = df["Rank"].apply(lambda r: MEDALS.get(r, str(r)))
|
83 |
df["Model"] = df.apply(lambda r: f"<a href='{r.url}' target='_blank'>{r.model_name}</a>", axis=1)
|
84 |
+
|
85 |
df.rename(columns={"organizer": "Organizer", "license": "License", col: "Elo Score"}, inplace=True)
|
86 |
return df[["Rank", "Model", "Organizer", "License", "Elo Score"]]
|
87 |
|
88 |
# ---------------------------------------------------------------------------
|
89 |
+
# CSS (dark + bigger fonts + heading icon)
|
90 |
# ---------------------------------------------------------------------------
|
91 |
custom_css += """
|
|
|
92 |
@import url('https://fonts.googleapis.com/css2?family=Inter:wght@400;500;600;700;800&display=swap');
|
93 |
+
html,body{font-family:'Inter',sans-serif!important;font-size:18px!important;line-height:1.6;color:#e2e8f0;background:#0f172a;}
|
|
|
|
|
|
|
|
|
|
|
|
|
94 |
|
95 |
+
.gradient-text{background:linear-gradient(90deg,#6366f1 0%,#06b6d4 100%);-webkit-background-clip:text;-webkit-text-fill-color:transparent;}
|
|
|
|
|
|
|
|
|
96 |
|
97 |
+
/* Section title */
|
98 |
+
.section-title{font-size:1.7rem;font-weight:700;color:#f1f5f9;display:flex;align-items:center;gap:.4em;margin:1.1em 0 .8em;}
|
99 |
+
.section-title .icon{font-size:1.4em;}
|
100 |
|
101 |
+
/* Radio & checkbox */
|
102 |
+
.gr-radio,.gr-checkbox{background:#1e293b;border-radius:8px;padding:6px 12px;box-shadow:0 1px 3px rgba(0,0,0,.4);}
|
103 |
+
.gr-radio input:checked+label,.gr-checkbox input:checked+label{color:#38bdf8;}
|
104 |
|
105 |
+
/* Leaderboard table */
|
106 |
#leaderboard-table table{width:100%;border-collapse:collapse;background:#1e293b;border-radius:8px;overflow:hidden;}
|
107 |
+
#leaderboard-table th{background:#334155;font-size:1rem;font-weight:600;padding:.8em;color:#f1f5f9;text-transform:uppercase;letter-spacing:.04em;}
|
108 |
+
#leaderboard-table td{padding:.7em;font-size:1.15rem;border-top:1px solid #334155;}
|
109 |
#leaderboard-table tr:nth-child(even){background:#1c2431;}
|
110 |
#leaderboard-table tr:hover{background:#475569;}
|
111 |
|
112 |
+
/* Scrollbar styling */
|
113 |
+
#leaderboard-table tbody{display:block;max-height:520px;overflow-y:auto;}
|
114 |
+
#leaderboard-table thead, #leaderboard-table tbody tr{display:table;width:100%;table-layout:fixed;}
|
115 |
+
|
116 |
/* Links */
|
117 |
a{color:#38bdf8;} a:hover{text-decoration:underline;}
|
118 |
|
|
|
119 |
.gr-accordion .label{font-weight:600;font-size:1rem;color:#f1f5f9;}
|
120 |
"""
|
121 |
|
|
|
130 |
|
131 |
with gr.Tabs():
|
132 |
with gr.TabItem("🏅 Leaderboard"):
|
133 |
+
gr.HTML("<h3 class='section-title'><span class='icon'>📊</span>Model Elo Rankings by Category</h3>")
|
134 |
with gr.Row():
|
135 |
category_radio = gr.Radio(CATEGORIES, value=DEFAULT_CATEGORY, label="Category")
|
136 |
+
asc_check = gr.Checkbox(label="⬆️ Asc. order", value=False)
|
137 |
+
|
138 |
board = gr.Dataframe(
|
139 |
value=update_leaderboard(DEFAULT_CATEGORY, False),
|
140 |
headers=["Rank", "Model", "Organizer", "License", "Elo Score"],
|
141 |
+
datatype=["html", "html", "str", "str", "number"],
|
142 |
row_count=(len(master_df), "fixed"),
|
143 |
col_count=(5, "fixed"),
|
144 |
interactive=False,
|
145 |
elem_id="leaderboard-table",
|
146 |
+
height=560,
|
147 |
)
|
148 |
category_radio.change(update_leaderboard, [category_radio, asc_check], board)
|
149 |
asc_check.change(update_leaderboard, [category_radio, asc_check], board)
|