Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
@@ -37,6 +37,84 @@ def sort_categories(categories):
|
|
37 |
# ))
|
38 |
# return templates
|
39 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
40 |
def get_modality_icon(modality):
|
41 |
"""Return an emoji icon for each modality type."""
|
42 |
icons = {
|
@@ -220,6 +298,9 @@ def update_detailed_scorecard(model, selected_categories):
|
|
220 |
if category_name in models[model]['scores']:
|
221 |
category_data = models[model]['scores'][category_name]
|
222 |
card_content = f"<div class='card'><div class='card-title'>{category_name}</div>"
|
|
|
|
|
|
|
223 |
|
224 |
# Sort sections within each category
|
225 |
sorted_sections = sorted(category_data.items(),
|
@@ -595,6 +676,153 @@ css = """
|
|
595 |
border-color: #2c3e50;
|
596 |
color: #99ccff;
|
597 |
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
598 |
"""
|
599 |
|
600 |
first_model = next(iter(models.values()))
|
|
|
37 |
# ))
|
38 |
# return templates
|
39 |
|
40 |
+
def create_category_summary(category_data):
|
41 |
+
"""Create a summary section for a category"""
|
42 |
+
# Calculate statistics
|
43 |
+
total_sections = len(category_data)
|
44 |
+
completed_sections = sum(1 for section in category_data.values() if section['status'] == 'Yes')
|
45 |
+
na_sections = sum(1 for section in category_data.values() if section['status'] == 'N/A')
|
46 |
+
|
47 |
+
# Calculate completion rates
|
48 |
+
total_questions = 0
|
49 |
+
completed_questions = 0
|
50 |
+
evaluation_types = set()
|
51 |
+
has_human_eval = False
|
52 |
+
has_quantitative = False
|
53 |
+
has_documentation = False
|
54 |
+
|
55 |
+
for section in category_data.values():
|
56 |
+
if section['status'] != 'N/A':
|
57 |
+
questions = section.get('questions', {})
|
58 |
+
total_questions += len(questions)
|
59 |
+
completed_questions += sum(1 for q in questions.values() if q)
|
60 |
+
|
61 |
+
# Check for evaluation types
|
62 |
+
for question in questions.keys():
|
63 |
+
if 'human' in question.lower():
|
64 |
+
has_human_eval = True
|
65 |
+
if any(term in question.lower() for term in ['quantitative', 'metric', 'benchmark']):
|
66 |
+
has_quantitative = True
|
67 |
+
if 'documentation' in question.lower():
|
68 |
+
has_documentation = True
|
69 |
+
|
70 |
+
completion_rate = (completed_questions / total_questions * 100) if total_questions > 0 else 0
|
71 |
+
|
72 |
+
# Create summary HTML
|
73 |
+
html = "<div class='summary-card'>"
|
74 |
+
html += "<div class='summary-title'>π Section Summary</div>"
|
75 |
+
|
76 |
+
# Completion metrics
|
77 |
+
html += "<div class='summary-section'>"
|
78 |
+
html += "<div class='summary-subtitle'>π Completion Metrics</div>"
|
79 |
+
html += f"<div class='metric-row'><span class='metric-label'>Overall Completion Rate:</span> <span class='metric-value'>{completion_rate:.1f}%</span></div>"
|
80 |
+
html += f"<div class='metric-row'><span class='metric-label'>Sections Completed:</span> <span class='metric-value'>{completed_sections}/{total_sections}</span></div>"
|
81 |
+
html += "</div>"
|
82 |
+
|
83 |
+
# Evaluation Coverage
|
84 |
+
html += "<div class='summary-section'>"
|
85 |
+
html += "<div class='summary-subtitle'>π― Evaluation Coverage</div>"
|
86 |
+
html += "<div class='coverage-grid'>"
|
87 |
+
html += f"<div class='coverage-item {get_coverage_class(has_human_eval)}'>π₯ Human Evaluation</div>"
|
88 |
+
html += f"<div class='coverage-item {get_coverage_class(has_quantitative)}'>π Quantitative Analysis</div>"
|
89 |
+
html += f"<div class='coverage-item {get_coverage_class(has_documentation)}'>π Documentation</div>"
|
90 |
+
html += "</div>"
|
91 |
+
html += "</div>"
|
92 |
+
|
93 |
+
# Status Breakdown
|
94 |
+
html += "<div class='summary-section'>"
|
95 |
+
html += "<div class='summary-subtitle'>π Status Breakdown</div>"
|
96 |
+
html += create_status_pills(category_data)
|
97 |
+
html += "</div>"
|
98 |
+
|
99 |
+
html += "</div>"
|
100 |
+
return html
|
101 |
+
|
102 |
+
def get_coverage_class(has_feature):
|
103 |
+
"""Return CSS class based on feature presence"""
|
104 |
+
return 'covered' if has_feature else 'not-covered'
|
105 |
+
|
106 |
+
def create_status_pills(category_data):
|
107 |
+
"""Create status pill indicators"""
|
108 |
+
status_counts = {'Yes': 0, 'No': 0, 'N/A': 0}
|
109 |
+
for section in category_data.values():
|
110 |
+
status_counts[section['status']] += 1
|
111 |
+
|
112 |
+
html = "<div class='status-pills'>"
|
113 |
+
for status, count in status_counts.items():
|
114 |
+
html += f"<div class='status-pill {status.lower()}'>{status}: {count}</div>"
|
115 |
+
html += "</div>"
|
116 |
+
return html
|
117 |
+
|
118 |
def get_modality_icon(modality):
|
119 |
"""Return an emoji icon for each modality type."""
|
120 |
icons = {
|
|
|
298 |
if category_name in models[model]['scores']:
|
299 |
category_data = models[model]['scores'][category_name]
|
300 |
card_content = f"<div class='card'><div class='card-title'>{category_name}</div>"
|
301 |
+
|
302 |
+
# Add summary section at the top of each card
|
303 |
+
card_content += create_category_summary(category_data)
|
304 |
|
305 |
# Sort sections within each category
|
306 |
sorted_sections = sorted(category_data.items(),
|
|
|
676 |
border-color: #2c3e50;
|
677 |
color: #99ccff;
|
678 |
}
|
679 |
+
|
680 |
+
.summary-card {
|
681 |
+
background-color: #f8f9fa;
|
682 |
+
border: 1px solid #e0e0e0;
|
683 |
+
border-radius: 8px;
|
684 |
+
padding: 16px;
|
685 |
+
margin-bottom: 20px;
|
686 |
+
}
|
687 |
+
|
688 |
+
.summary-title {
|
689 |
+
font-size: 1.2em;
|
690 |
+
font-weight: bold;
|
691 |
+
margin-bottom: 12px;
|
692 |
+
color: #333;
|
693 |
+
}
|
694 |
+
|
695 |
+
.summary-section {
|
696 |
+
margin-bottom: 16px;
|
697 |
+
}
|
698 |
+
|
699 |
+
.summary-subtitle {
|
700 |
+
font-size: 1em;
|
701 |
+
font-weight: 600;
|
702 |
+
color: #555;
|
703 |
+
margin-bottom: 8px;
|
704 |
+
}
|
705 |
+
|
706 |
+
.metric-row {
|
707 |
+
display: flex;
|
708 |
+
justify-content: space-between;
|
709 |
+
align-items: center;
|
710 |
+
margin-bottom: 4px;
|
711 |
+
}
|
712 |
+
|
713 |
+
.metric-label {
|
714 |
+
color: #666;
|
715 |
+
}
|
716 |
+
|
717 |
+
.metric-value {
|
718 |
+
font-weight: 600;
|
719 |
+
color: #333;
|
720 |
+
}
|
721 |
+
|
722 |
+
.coverage-grid {
|
723 |
+
display: grid;
|
724 |
+
grid-template-columns: repeat(auto-fit, minmax(150px, 1fr));
|
725 |
+
gap: 8px;
|
726 |
+
margin-top: 8px;
|
727 |
+
}
|
728 |
+
|
729 |
+
.coverage-item {
|
730 |
+
padding: 8px;
|
731 |
+
border-radius: 6px;
|
732 |
+
text-align: center;
|
733 |
+
font-size: 0.9em;
|
734 |
+
}
|
735 |
+
|
736 |
+
.coverage-item.covered {
|
737 |
+
background-color: #e6ffe6;
|
738 |
+
color: #006600;
|
739 |
+
border: 1px solid #b3ffb3;
|
740 |
+
}
|
741 |
+
|
742 |
+
.coverage-item.not-covered {
|
743 |
+
background-color: #f5f5f5;
|
744 |
+
color: #666;
|
745 |
+
border: 1px solid #ddd;
|
746 |
+
}
|
747 |
+
|
748 |
+
.status-pills {
|
749 |
+
display: flex;
|
750 |
+
gap: 8px;
|
751 |
+
flex-wrap: wrap;
|
752 |
+
}
|
753 |
+
|
754 |
+
.status-pill {
|
755 |
+
padding: 4px 12px;
|
756 |
+
border-radius: 16px;
|
757 |
+
font-size: 0.9em;
|
758 |
+
font-weight: 500;
|
759 |
+
}
|
760 |
+
|
761 |
+
.status-pill.yes {
|
762 |
+
background-color: #e6ffe6;
|
763 |
+
color: #006600;
|
764 |
+
border: 1px solid #b3ffb3;
|
765 |
+
}
|
766 |
+
|
767 |
+
.status-pill.no {
|
768 |
+
background-color: #ffe6e6;
|
769 |
+
color: #990000;
|
770 |
+
border: 1px solid #ffb3b3;
|
771 |
+
}
|
772 |
+
|
773 |
+
.status-pill.n\\/a {
|
774 |
+
background-color: #f5f5f5;
|
775 |
+
color: #666;
|
776 |
+
border: 1px solid #ddd;
|
777 |
+
}
|
778 |
+
|
779 |
+
.dark .summary-card {
|
780 |
+
background-color: #2a2a2a;
|
781 |
+
border-color: #444;
|
782 |
+
}
|
783 |
+
|
784 |
+
.dark .summary-title,
|
785 |
+
.dark .summary-subtitle {
|
786 |
+
color: #e0e0e0;
|
787 |
+
}
|
788 |
+
|
789 |
+
.dark .metric-label {
|
790 |
+
color: #999;
|
791 |
+
}
|
792 |
+
|
793 |
+
.dark .metric-value {
|
794 |
+
color: #fff;
|
795 |
+
}
|
796 |
+
|
797 |
+
.dark .coverage-item.covered {
|
798 |
+
background-color: #1a3a1a;
|
799 |
+
color: #90EE90;
|
800 |
+
border-color: #2d5a2d;
|
801 |
+
}
|
802 |
+
|
803 |
+
.dark .coverage-item.not-covered {
|
804 |
+
background-color: #333;
|
805 |
+
color: #999;
|
806 |
+
border-color: #444;
|
807 |
+
}
|
808 |
+
|
809 |
+
.dark .status-pill.yes {
|
810 |
+
background-color: #1a3a1a;
|
811 |
+
color: #90EE90;
|
812 |
+
border-color: #2d5a2d;
|
813 |
+
}
|
814 |
+
|
815 |
+
.dark .status-pill.no {
|
816 |
+
background-color: #3a1a1a;
|
817 |
+
color: #FFB6B6;
|
818 |
+
border-color: #5a2d2d;
|
819 |
+
}
|
820 |
+
|
821 |
+
.dark .status-pill.n\\/a {
|
822 |
+
background-color: #333;
|
823 |
+
color: #999;
|
824 |
+
border-color: #444;
|
825 |
+
}
|
826 |
"""
|
827 |
|
828 |
first_model = next(iter(models.values()))
|