Spaces:
Running
Running
Update app.py
Browse files
app.py
CHANGED
@@ -3,17 +3,14 @@ import pandas as pd
|
|
3 |
from apscheduler.schedulers.background import BackgroundScheduler
|
4 |
|
5 |
"""
|
6 |
-
MLE‑Dojo Benchmark Leaderboard — Dark Elegance
|
7 |
================================================
|
8 |
-
|
9 |
-
*
|
10 |
-
* larger fonts throughout
|
11 |
-
* medal icons for top‑3 ranks
|
12 |
-
* scrollable leaderboard (fixed height)
|
13 |
"""
|
14 |
|
15 |
# ---------------------------------------------------------------------------
|
16 |
-
# Import
|
17 |
# ---------------------------------------------------------------------------
|
18 |
try:
|
19 |
from src.about import (
|
@@ -23,7 +20,7 @@ try:
|
|
23 |
LLM_BENCHMARKS_TEXT,
|
24 |
TITLE,
|
25 |
)
|
26 |
-
from src.display.css_html_js import custom_css
|
27 |
from src.envs import REPO_ID
|
28 |
from src.submission.submit import add_new_eval
|
29 |
except ImportError:
|
@@ -43,6 +40,8 @@ except ImportError:
|
|
43 |
# ---------------------------------------------------------------------------
|
44 |
# Data
|
45 |
# ---------------------------------------------------------------------------
|
|
|
|
|
46 |
data = [
|
47 |
{"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},
|
48 |
{"model_name": "gpt-4o", "url": "https://openai.com/index/hello-gpt-4o/", "organizer": "OpenAI", "license": "Proprietary", "MLE-Lite_Elo": 830, "Tabular_Elo": 861, "NLP_Elo": 903, "CV_Elo": 761, "Overall": 841},
|
@@ -67,7 +66,6 @@ CATEGORY_MAP = {
|
|
67 |
"NLP": "NLP_Elo",
|
68 |
"CV": "CV_Elo",
|
69 |
}
|
70 |
-
|
71 |
MEDALS = {1: "🥇", 2: "🥈", 3: "🥉"}
|
72 |
|
73 |
def update_leaderboard(category: str, ascending: bool):
|
@@ -77,46 +75,17 @@ def update_leaderboard(category: str, ascending: bool):
|
|
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 +
|
90 |
# ---------------------------------------------------------------------------
|
91 |
custom_css += """
|
92 |
-
|
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 |
|
122 |
# ---------------------------------------------------------------------------
|
@@ -127,14 +96,12 @@ app = gr.Blocks(css=custom_css, theme=gr.themes.Soft(primary_hue="sky", neutral_
|
|
127 |
with app:
|
128 |
gr.HTML(TITLE)
|
129 |
gr.Markdown(INTRODUCTION_TEXT, elem_classes="markdown-text")
|
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"],
|
@@ -143,19 +110,16 @@ with app:
|
|
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)
|
150 |
-
|
151 |
-
with gr.TabItem("ℹ️ About"):
|
152 |
gr.Markdown(LLM_BENCHMARKS_TEXT, elem_classes="markdown-text")
|
153 |
-
|
154 |
-
with gr.Accordion("📖 Citation", open=False):
|
155 |
gr.Textbox(value=CITATION_BUTTON_TEXT, label=CITATION_BUTTON_LABEL, lines=10, show_copy_button=True)
|
156 |
|
157 |
# ---------------------------------------------------------------------------
|
158 |
-
# Optional scheduler
|
159 |
# ---------------------------------------------------------------------------
|
160 |
|
161 |
def restart_space():
|
|
|
3 |
from apscheduler.schedulers.background import BackgroundScheduler
|
4 |
|
5 |
"""
|
6 |
+
MLE‑Dojo Benchmark Leaderboard — Dark Elegance v3
|
7 |
================================================
|
8 |
+
* Fix: removed unsupported `height` param for `gr.Dataframe`.
|
9 |
+
* Font tweak: leaderboard cells slightly smaller.
|
|
|
|
|
|
|
10 |
"""
|
11 |
|
12 |
# ---------------------------------------------------------------------------
|
13 |
+
# Import copy or fallback placeholders
|
14 |
# ---------------------------------------------------------------------------
|
15 |
try:
|
16 |
from src.about import (
|
|
|
20 |
LLM_BENCHMARKS_TEXT,
|
21 |
TITLE,
|
22 |
)
|
23 |
+
from src.display.css_html_js import custom_css
|
24 |
from src.envs import REPO_ID
|
25 |
from src.submission.submit import add_new_eval
|
26 |
except ImportError:
|
|
|
40 |
# ---------------------------------------------------------------------------
|
41 |
# Data
|
42 |
# ---------------------------------------------------------------------------
|
43 |
+
# (unchanged)-------------------------------------------
|
44 |
+
|
45 |
data = [
|
46 |
{"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},
|
47 |
{"model_name": "gpt-4o", "url": "https://openai.com/index/hello-gpt-4o/", "organizer": "OpenAI", "license": "Proprietary", "MLE-Lite_Elo": 830, "Tabular_Elo": 861, "NLP_Elo": 903, "CV_Elo": 761, "Overall": 841},
|
|
|
66 |
"NLP": "NLP_Elo",
|
67 |
"CV": "CV_Elo",
|
68 |
}
|
|
|
69 |
MEDALS = {1: "🥇", 2: "🥈", 3: "🥉"}
|
70 |
|
71 |
def update_leaderboard(category: str, ascending: bool):
|
|
|
75 |
.sort_values(by=col, ascending=ascending)
|
76 |
.reset_index(drop=True)
|
77 |
)
|
|
|
78 |
df.insert(0, "Rank", df.index + 1)
|
79 |
df["Rank"] = df["Rank"].apply(lambda r: MEDALS.get(r, str(r)))
|
80 |
df["Model"] = df.apply(lambda r: f"<a href='{r.url}' target='_blank'>{r.model_name}</a>", axis=1)
|
|
|
81 |
df.rename(columns={"organizer": "Organizer", "license": "License", col: "Elo Score"}, inplace=True)
|
82 |
return df[["Rank", "Model", "Organizer", "License", "Elo Score"]]
|
83 |
|
84 |
# ---------------------------------------------------------------------------
|
85 |
+
# CSS (dark + slightly smaller table font)
|
86 |
# ---------------------------------------------------------------------------
|
87 |
custom_css += """
|
88 |
+
#leaderboard-table td{padding:.7em;font-size:1.05rem;border-top:1px solid #334155;}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
89 |
"""
|
90 |
|
91 |
# ---------------------------------------------------------------------------
|
|
|
96 |
with app:
|
97 |
gr.HTML(TITLE)
|
98 |
gr.Markdown(INTRODUCTION_TEXT, elem_classes="markdown-text")
|
|
|
99 |
with gr.Tabs():
|
100 |
with gr.TabItem("🏅 Leaderboard"):
|
101 |
gr.HTML("<h3 class='section-title'><span class='icon'>📊</span>Model Elo Rankings by Category</h3>")
|
102 |
with gr.Row():
|
103 |
category_radio = gr.Radio(CATEGORIES, value=DEFAULT_CATEGORY, label="Category")
|
104 |
asc_check = gr.Checkbox(label="⬆️ Asc. order", value=False)
|
|
|
105 |
board = gr.Dataframe(
|
106 |
value=update_leaderboard(DEFAULT_CATEGORY, False),
|
107 |
headers=["Rank", "Model", "Organizer", "License", "Elo Score"],
|
|
|
110 |
col_count=(5, "fixed"),
|
111 |
interactive=False,
|
112 |
elem_id="leaderboard-table",
|
|
|
113 |
)
|
114 |
category_radio.change(update_leaderboard, [category_radio, asc_check], board)
|
115 |
asc_check.change(update_leaderboard, [category_radio, asc_check], board)
|
116 |
+
with gr.TabItem("ℹ️ About"):
|
|
|
117 |
gr.Markdown(LLM_BENCHMARKS_TEXT, elem_classes="markdown-text")
|
118 |
+
with gr.Accordion("📖 Citation", open=False):
|
|
|
119 |
gr.Textbox(value=CITATION_BUTTON_TEXT, label=CITATION_BUTTON_LABEL, lines=10, show_copy_button=True)
|
120 |
|
121 |
# ---------------------------------------------------------------------------
|
122 |
+
# Optional scheduler
|
123 |
# ---------------------------------------------------------------------------
|
124 |
|
125 |
def restart_space():
|