Delete app.py
Browse files
app.py
DELETED
@@ -1,115 +0,0 @@
|
|
1 |
-
"""
|
2 |
-
Main application for Dynamic Highscores system.
|
3 |
-
|
4 |
-
This file integrates all components into a unified application.
|
5 |
-
"""
|
6 |
-
|
7 |
-
import os
|
8 |
-
import gradio as gr
|
9 |
-
import threading
|
10 |
-
import queue
|
11 |
-
from database_schema import init_db
|
12 |
-
from auth import HuggingFaceAuth, create_login_ui, setup_auth_handlers
|
13 |
-
from benchmark_selection import BenchmarkSelector, create_benchmark_selection_ui
|
14 |
-
from evaluation_queue import EvaluationQueue, create_model_submission_ui
|
15 |
-
from leaderboard import Leaderboard, create_leaderboard_ui
|
16 |
-
from sample_benchmarks import add_sample_benchmarks
|
17 |
-
|
18 |
-
# Initialize database
|
19 |
-
db = init_db()
|
20 |
-
|
21 |
-
# Initialize components
|
22 |
-
auth_manager = HuggingFaceAuth(db)
|
23 |
-
benchmark_selector = BenchmarkSelector(db, auth_manager)
|
24 |
-
evaluation_queue = EvaluationQueue(db, auth_manager)
|
25 |
-
leaderboard = Leaderboard(db)
|
26 |
-
|
27 |
-
# Initialize sample benchmarks if none exist
|
28 |
-
benchmarks = db.get_benchmarks()
|
29 |
-
if not benchmarks or len(benchmarks) == 0:
|
30 |
-
print("No benchmarks found. Adding sample benchmarks...")
|
31 |
-
num_added = add_sample_benchmarks()
|
32 |
-
print(f"Added {num_added} sample benchmarks.")
|
33 |
-
|
34 |
-
# Custom CSS
|
35 |
-
css = """
|
36 |
-
.info-text {
|
37 |
-
background-color: #f0f7ff;
|
38 |
-
padding: 12px;
|
39 |
-
border-radius: 8px;
|
40 |
-
border-left: 4px solid #3498db;
|
41 |
-
margin: 12px 0;
|
42 |
-
}
|
43 |
-
|
44 |
-
.container {
|
45 |
-
max-width: 1200px;
|
46 |
-
margin: 0 auto;
|
47 |
-
}
|
48 |
-
|
49 |
-
.header {
|
50 |
-
text-align: center;
|
51 |
-
margin-bottom: 20px;
|
52 |
-
}
|
53 |
-
|
54 |
-
.footer {
|
55 |
-
text-align: center;
|
56 |
-
margin-top: 40px;
|
57 |
-
padding: 20px;
|
58 |
-
border-top: 1px solid #eee;
|
59 |
-
}
|
60 |
-
"""
|
61 |
-
|
62 |
-
# Create Gradio app
|
63 |
-
with gr.Blocks(css=css, title="Dynamic Highscores") as app:
|
64 |
-
gr.Markdown("# 🏆 Dynamic Highscores", elem_classes=["header"])
|
65 |
-
gr.Markdown("""
|
66 |
-
Welcome to Dynamic Highscores - a community benchmark platform for evaluating and comparing language models.
|
67 |
-
|
68 |
-
- **Add your own benchmarks** from HuggingFace datasets
|
69 |
-
- **Submit your models** for CPU-only evaluation
|
70 |
-
- **Compare performance** across different models and benchmarks
|
71 |
-
- **Filter results** by model type (Merge, Agent, Reasoning, Coding, etc.)
|
72 |
-
""", elem_classes=["info-text"])
|
73 |
-
|
74 |
-
# Authentication UI
|
75 |
-
login_button, logout_button, token_input, user_info = create_login_ui()
|
76 |
-
setup_auth_handlers(login_button, logout_button, token_input, user_info, auth_manager)
|
77 |
-
|
78 |
-
# Main tabs
|
79 |
-
with gr.Tabs() as tabs:
|
80 |
-
with gr.TabItem("📊 Leaderboard", id=0):
|
81 |
-
# Fix: Pass db_manager parameter to create_leaderboard_ui
|
82 |
-
leaderboard_ui = create_leaderboard_ui(leaderboard, db)
|
83 |
-
|
84 |
-
with gr.TabItem("🚀 Submit Model", id=1):
|
85 |
-
submission_ui = create_model_submission_ui(evaluation_queue, auth_manager, db)
|
86 |
-
|
87 |
-
with gr.TabItem("🔍 Benchmarks", id=2):
|
88 |
-
benchmark_ui = create_benchmark_selection_ui(benchmark_selector, auth_manager)
|
89 |
-
|
90 |
-
gr.Markdown("""
|
91 |
-
### About Dynamic Highscores
|
92 |
-
|
93 |
-
This platform allows users to select benchmarks from HuggingFace datasets and evaluate models against them.
|
94 |
-
Each user can submit one benchmark per day (admin users are exempt from this limit).
|
95 |
-
All evaluations run on CPU only to ensure fair comparisons.
|
96 |
-
|
97 |
-
Created by Quazim0t0
|
98 |
-
""", elem_classes=["footer"])
|
99 |
-
|
100 |
-
# Start evaluation queue worker after app is defined
|
101 |
-
# This prevents the worker from starting before the app is fully initialized
|
102 |
-
def start_queue_worker():
|
103 |
-
# Wait a moment to ensure app is initialized
|
104 |
-
import time
|
105 |
-
time.sleep(2)
|
106 |
-
evaluation_queue.start_worker()
|
107 |
-
|
108 |
-
# Launch the app
|
109 |
-
if __name__ == "__main__":
|
110 |
-
# Start queue worker in a separate thread to avoid SQLite thread issues
|
111 |
-
queue_thread = threading.Thread(target=start_queue_worker)
|
112 |
-
queue_thread.daemon = True
|
113 |
-
queue_thread.start()
|
114 |
-
|
115 |
-
app.launch()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|