Spaces:
Running
Running
File size: 45,479 Bytes
53a73e0 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 |
"""
PazaBench Visualization Functions for Gradio Integration
"""
import os
import pandas as pd
import plotly.graph_objects as go
from src.constants import (
COUNTRY_NAMES,
LANGUAGE_COUNTRY_MAP,
LANGUAGE_SAMPLE_COUNTS,
)
from src.model_counts import MODEL_PARAMETER_COUNTS
# Load model family colors from CSV
def _load_model_family_colors() -> tuple[dict[str, str], dict[str, str]]:
"""
Load color mappings from the model_family_colors.csv file.
Returns:
- model_family_colors: dict mapping model_family -> color (first color for that family)
- model_id_colors: dict mapping model_id -> color
"""
csv_path = os.path.join(os.path.dirname(__file__), 'display', 'model_family_colors.csv')
model_family_colors = {}
model_id_colors = {}
try:
df = pd.read_csv(csv_path)
for _, row in df.iterrows():
family = row['Model Families']
model_id = row['Model ID']
color = row['Color']
# Store first color encountered for each family (using normalized key)
normalized_family = _normalize_family_name(family)
if normalized_family not in model_family_colors:
model_family_colors[normalized_family] = color
# Store color for each model ID
model_id_colors[model_id] = color
except Exception as e:
print(f"Warning: Could not load model family colors: {e}")
return model_family_colors, model_id_colors
def _normalize_family_name(family: str) -> str:
"""Normalize family name for consistent lookup (lowercase, remove underscores/dashes)."""
return family.lower().replace('_', '').replace('-', '').replace(' ', '')
MODEL_FAMILY_COLORS, MODEL_ID_COLORS = _load_model_family_colors()
def _get_color_for_family(family: str) -> str:
"""Get color for a model family, with fallback."""
normalized = _normalize_family_name(family)
return MODEL_FAMILY_COLORS.get(normalized, '#888888')
def _get_color_for_model(model_id: str) -> str:
"""Get color for a specific model ID, with fallback to family color."""
if model_id in MODEL_ID_COLORS:
return MODEL_ID_COLORS[model_id]
# Try to find family and use family color
for family, color in MODEL_FAMILY_COLORS.items():
if family.lower() in model_id.lower():
return color
return '#888888'
def _remove_wer_outliers(df: pd.DataFrame, multiplier: float = 1.5) -> pd.DataFrame:
"""
Remove WER outliers using IQR method for cleaner visualizations.
Only removes HIGH outliers (poor performers), keeps LOW outliers (best performers).
Args:
df: DataFrame with 'wer' column
multiplier: IQR multiplier (default 1.5 for standard outlier detection)
Returns:
DataFrame with high outliers removed
"""
if df.empty or 'wer' not in df.columns:
return df
Q1 = df['wer'].quantile(0.25)
Q3 = df['wer'].quantile(0.75)
IQR = Q3 - Q1
# Only remove HIGH outliers (poor performers), keep LOW outliers (best performers)
# Lower WER is better, so we don't want to remove low-WER entries
upper_bound = Q3 + multiplier * IQR
return df[df['wer'] <= upper_bound]
def create_language_coverage_chart(selected_languages: list[str] | None = None) -> go.Figure:
"""
Create a horizontal bar chart showing sample counts for each language in PazaBench.
Languages are sorted by sample count (descending).
Selected languages are highlighted with a different color.
Args:
selected_languages: List of languages to highlight (None = no highlighting)
"""
# Create dataframe from language sample counts
df = pd.DataFrame([
{"Language": lang, "Sample Count": count}
for lang, count in LANGUAGE_SAMPLE_COUNTS.items()
])
# Sort by sample count descending
df = df.sort_values("Sample Count", ascending=True) # ascending=True for horizontal bar (bottom to top)
# Determine colors - highlight selected languages
if selected_languages and len(selected_languages) > 0:
colors = [
"#0078D4" if lang in selected_languages else "#D0E8FF"
for lang in df["Language"]
]
# Add border to selected bars
line_widths = [2 if lang in selected_languages else 0 for lang in df["Language"]]
line_colors = ["#005A9E" if lang in selected_languages else "rgba(0,0,0,0)" for lang in df["Language"]]
else:
colors = "#8CD0FF" # Solid blue color matching theme
line_widths = 0
line_colors = "rgba(0,0,0,0)"
fig = go.Figure(go.Bar(
y=df["Language"],
x=df["Sample Count"],
orientation='h',
marker=dict(
color=colors,
line=dict(width=line_widths, color=line_colors) if selected_languages else None,
),
text=None, # Remove text labels above bars
textposition='none',
hovertemplate='<b>%{y}</b><br>Samples: %{x:,}<extra></extra>',
))
# Build title
if selected_languages and len(selected_languages) > 0:
title_text = f"Language Coverage in PazaBench ({len(selected_languages)} selected)"
else:
title_text = "Language Coverage in PazaBench"
fig.update_layout(
title=dict(
text=title_text,
font=dict(size=16),
x=0.5
),
xaxis_title="Number of Samples",
yaxis_title="",
height=800,
autosize=True,
margin=dict(l=120, r=50, t=60, b=40),
template='plotly_white',
showlegend=False,
)
return fig
def create_language_location_map(languages: str | list[str] | None = None) -> go.Figure:
"""
Create an interactive choropleth map of Africa showing where specific language(s) exist.
If no language is selected, shows a light overview map with all PazaBench countries highlighted.
Args:
languages: The language(s) to highlight on the map (single string or list)
"""
fig = go.Figure()
# Normalize to list
if isinstance(languages, str):
languages = [languages]
if languages and len(languages) > 0:
# Get countries where these languages exist
all_countries = set()
country_language_map = {} # Track which languages are spoken in each country
for lang in languages:
if lang in LANGUAGE_COUNTRY_MAP:
for code in LANGUAGE_COUNTRY_MAP[lang]:
all_countries.add(code)
if code not in country_language_map:
country_language_map[code] = []
country_language_map[code].append(lang)
if all_countries:
# Create dataframe with countries that have these languages
df_map = pd.DataFrame([
{
"country_code": code,
"country_name": COUNTRY_NAMES.get(code, code),
"has_language": 1,
"languages": ", ".join(country_language_map.get(code, []))
}
for code in all_countries
])
# Build hover text
if len(languages) == 1:
hover_template = "<b>%{text}</b><br>" + f"{languages[0]} is spoken here<extra></extra>"
else:
hover_template = "<b>%{text}</b><br>Languages: %{customdata}<extra></extra>"
fig.add_trace(go.Choropleth(
locations=df_map["country_code"],
z=df_map["has_language"],
text=df_map["country_name"],
customdata=df_map["languages"],
hovertemplate=hover_template,
colorscale=[[0, "#8CD0FF"], [1, "#8CD0FF"]], # Solid blue color
showscale=False,
marker_line_color="white",
marker_line_width=0.5,
))
if len(languages) == 1:
title_text = f"Where {languages[0]} is Spoken"
else:
title_text = f"Where {len(languages)} Selected Languages are Spoken"
else:
title_text = "Select a Language to Explore"
else:
# Default view: show all countries with PazaBench data lightly highlighted
all_countries = set()
for countries in LANGUAGE_COUNTRY_MAP.values():
all_countries.update(countries)
df_map = pd.DataFrame([
{
"country_code": code,
"country_name": COUNTRY_NAMES.get(code, code),
"in_pazabench": 1
}
for code in all_countries
])
fig.add_trace(go.Choropleth(
locations=df_map["country_code"],
z=df_map["in_pazabench"],
text=df_map["country_name"],
hovertemplate="<b>%{text}</b><br>Has PazaBench data<extra></extra>",
colorscale=[[0, "#E8F4FF"], [1, "#E8F4FF"]], # Very light blue matching theme
showscale=False,
marker_line_color="#B3D9FF",
marker_line_width=0.5,
))
title_text = "Select a Language to Explore"
fig.update_geos(
visible=True,
resolution=50,
scope="africa",
showcountries=True,
countrycolor="lightgray",
showcoastlines=True,
coastlinecolor="gray",
showland=True,
landcolor="#f5f5f5",
showocean=True,
oceancolor="#e3f2fd",
showlakes=True,
lakecolor="#e3f2fd",
projection_type="natural earth",
center=dict(lat=5, lon=20),
)
fig.update_layout(
title=dict(
text=title_text,
font=dict(size=18),
x=0.5
),
height=600,
autosize=True,
margin=dict(l=5, r=5, t=50, b=5),
geo=dict(
bgcolor="rgba(0,0,0,0)",
)
)
# Use SVG renderer for better resolution
fig.update_layout(
template="plotly_white",
)
return fig
def get_language_sample_info(languages: str | list[str] | None = None, asr_df: pd.DataFrame | None = None) -> str:
"""
Get the sample count information for specific language(s) as styled HTML.
Returns HTML for display in Gradio.
Args:
languages: The language(s) to get information for (single string or list)
asr_df: DataFrame with ASR results to extract dataset groups
"""
# Normalize to list
if isinstance(languages, str):
languages = [languages]
if languages and len(languages) > 0:
# Filter to valid languages
valid_languages = [lang for lang in languages if lang in LANGUAGE_SAMPLE_COUNTS]
if valid_languages:
# Aggregate data across all selected languages
total_samples = sum(LANGUAGE_SAMPLE_COUNTS.get(lang, 0) for lang in valid_languages)
all_countries = set()
for lang in valid_languages:
all_countries.update(LANGUAGE_COUNTRY_MAP.get(lang, []))
country_names = sorted([COUNTRY_NAMES.get(code, code) for code in all_countries])
# Get dataset groups from ASR data if available
dataset_groups = set()
if asr_df is not None and not asr_df.empty and 'language' in asr_df.columns:
for lang in valid_languages:
lang_data = asr_df[asr_df['language'] == lang]
if not lang_data.empty and 'dataset_group' in lang_data.columns:
dataset_groups.update(lang_data['dataset_group'].unique().tolist())
dataset_groups = sorted(dataset_groups)
# Build title based on number of languages
if len(valid_languages) == 1:
title = f"🌍 {valid_languages[0]}"
else:
title = f"🌍 {', '.join(sorted(valid_languages))}"
html = f"""
<div style="background: linear-gradient(135deg, #DDF1FF 0%, #ffffff 100%); border-radius: 12px; padding: 20px; border: 1px solid #8CD0FF;">
<h4 style="margin: 0 0 16px 0; color: #0f172a; font-size: 1.1em; font-weight: 600; border-bottom: 2px solid #8CD0FF; padding-bottom: 8px;">
{title}
</h4>
<div style="display: grid; gap: 12px;">
<div style="background: white; padding: 12px 16px; border-radius: 8px; display: flex; justify-content: space-between; align-items: center;">
<span style="color: #666; font-weight: 500; font-size: 0.9rem;">📊 Total Samples</span>
<span style="color: #0f172a; font-weight: 700; font-size: 0.9rem;">{total_samples:,}</span>
</div>
<div style="background: white; padding: 12px 16px; border-radius: 8px;">
<div style="color: #666; font-weight: 500; margin-bottom: 6px; font-size: 0.9rem;">📍 Countries ({len(country_names)})</div>
<div style="color: #0f172a; font-weight: 500; font-size: 0.9rem;">{', '.join(country_names) if country_names else 'N/A'}</div>
</div>
<div style="background: white; padding: 12px 16px; border-radius: 8px;">
<div style="color: #666; font-weight: 500; margin-bottom: 6px; font-size: 0.9rem;">📁 Dataset Sources ({len(dataset_groups)})</div>
<div style="color: #0f172a; font-size: 0.9rem;">
{(', '.join(dataset_groups) if dataset_groups else '<em>No dataset info available</em>')}
</div>
</div>
</div>
</div>
"""
return html
# Default view - show sample overview summary
total_languages = len(LANGUAGE_SAMPLE_COUNTS)
total_samples = sum(LANGUAGE_SAMPLE_COUNTS.values())
total_countries = len(set(code for codes in LANGUAGE_COUNTRY_MAP.values() for code in codes))
html = f"""
<div style="background: linear-gradient(135deg, #f0f9ff 0%, #ffffff 100%); border-radius: 12px; padding: 20px; border: 1px solid #e0e7ef;">
<h4 style="margin: 0 0 16px 0; color: #0f172a; font-size: 1.1em; font-weight: 600; border-bottom: 2px solid #8CD0FF; padding-bottom: 8px;">
📊 Sample Overview
</h4>
<div style="display: grid; gap: 10px; margin-bottom: 16px;">
<div style="background: white; padding: 10px 14px; border-radius: 8px; display: flex; justify-content: space-between; align-items: center; border: 1px solid #e5e7eb;">
<span style="color: #666; font-weight: 500; font-size: 0.9rem;">🌍 Languages</span>
<span style="color: #0f172a; font-weight: 700; font-size: 0.9rem;">{total_languages}</span>
</div>
<div style="background: white; padding: 10px 14px; border-radius: 8px; display: flex; justify-content: space-between; align-items: center; border: 1px solid #e5e7eb;">
<span style="color: #666; font-weight: 500; font-size: 0.9rem;">📊 Total Samples</span>
<span style="color: #0f172a; font-weight: 700; font-size: 0.9rem;">{total_samples:,}</span>
</div>
<div style="background: white; padding: 10px 14px; border-radius: 8px; display: flex; justify-content: space-between; align-items: center; border: 1px solid #e5e7eb;">
<span style="color: #666; font-weight: 500; font-size: 0.9rem;">📍 Countries</span>
<span style="color: #0f172a; font-weight: 700; font-size: 0.9rem;">{total_countries}</span>
</div>
</div>
<p style="margin: 14px 0 0 0; font-size: 0.85rem; color: #666; font-style: italic; text-align: center;">
👈 Select a language on the left to explore its details
</p>
</div>
"""
return html
def get_language_sample_info_df(language: str | None = None) -> pd.DataFrame:
"""
Legacy function - returns DataFrame for backward compatibility.
"""
if language and language in LANGUAGE_SAMPLE_COUNTS:
sample_count = LANGUAGE_SAMPLE_COUNTS[language]
countries = LANGUAGE_COUNTRY_MAP.get(language, [])
country_names = [COUNTRY_NAMES.get(code, code) for code in countries]
df = pd.DataFrame({
"Metric": ["Language", "Total Samples", "Countries"],
"Value": [
language,
f"{sample_count:,}",
", ".join(country_names) if country_names else "N/A"
]
})
else:
df = pd.DataFrame({
"Metric": ["Language", "Total Samples", "Countries"],
"Value": ["Select a language", "-", "-"]
})
return df
def get_all_languages() -> list[str]:
"""Get a sorted list of all languages in PazaBench."""
return sorted(LANGUAGE_SAMPLE_COUNTS.keys())
def create_africa_language_map() -> go.Figure:
"""
Create an interactive choropleth map of Africa showing language coverage.
Hover over countries to see the languages spoken there.
"""
# Build country data with languages
country_data = {}
for language, countries in LANGUAGE_COUNTRY_MAP.items():
for country_code in countries:
if country_code not in country_data:
country_data[country_code] = {
"languages": [],
"count": 0,
"country_name": COUNTRY_NAMES.get(country_code, country_code)
}
country_data[country_code]["languages"].append(language)
country_data[country_code]["count"] += 1
# Create dataframe for plotly
df_map = pd.DataFrame([
{
"country_code": code,
"country_name": data["country_name"],
"language_count": data["count"],
"languages": ", ".join(sorted(data["languages"]))
}
for code, data in country_data.items()
])
fig = go.Figure(go.Choropleth(
locations=df_map["country_code"],
z=df_map["language_count"],
text=df_map["country_name"],
customdata=df_map[["languages", "language_count"]],
hovertemplate="<b>%{text}</b><br>" +
"Languages: %{customdata[1]}<br>" +
"<i>%{customdata[0]}</i><extra></extra>",
colorscale=[
[0, "#DDF1FF"],
[0.25, "#B5E0FF"],
[0.5, "#8CD0FF"],
[0.75, "#6BC0F5"],
[1, "#4AAFEB"]
],
showscale=True,
colorbar=dict(
title="Languages",
tickmode="linear",
tick0=1,
dtick=1
),
marker_line_color="white",
marker_line_width=0.5,
))
fig.update_geos(
visible=True,
resolution=50,
scope="africa",
showcountries=True,
countrycolor="lightgray",
showcoastlines=True,
coastlinecolor="gray",
showland=True,
landcolor="#f5f5f5",
showocean=True,
oceancolor="#e3f2fd",
showlakes=True,
lakecolor="#e3f2fd",
projection_type="natural earth",
center=dict(lat=5, lon=20),
)
fig.update_layout(
title=dict(
text="African Languages in PazaBench",
font=dict(size=18),
x=0.5
),
height=600,
autosize=True,
margin=dict(l=5, r=5, t=50, b=5),
geo=dict(
bgcolor="rgba(0,0,0,0)",
),
template="plotly_white",
)
return fig
def create_model_leaderboard(df: pd.DataFrame, languages: list[str] | None = None, top_n_models: int = 15) -> go.Figure:
"""
Visualization 1: Model Family / Individual Model Performance Leaderboard
- When no language filter: Shows model families (aggregated)
- When language(s) selected: Shows top N individual models
Outliers are removed for cleaner visualization.
Args:
df: DataFrame with evaluation results
languages: List of languages to filter by (None = all languages)
top_n_models: Number of top individual models to show when languages are filtered (default: 15)
"""
# Apply language filter if provided
filtered_df = df.copy()
if languages:
filtered_df = filtered_df[filtered_df['language'].isin(languages)]
# Remove WER outliers for cleaner visualization
filtered_df = _remove_wer_outliers(filtered_df)
# Determine mode: individual models if languages selected, otherwise model families
show_individual_models = languages is not None and len(languages) > 0
if show_individual_models:
# Individual model mode: show top N models by median WER
model_perf = filtered_df.groupby(['model_family', 'model']).agg({
'wer': ['median', 'std', 'count'],
'cer': 'median',
'rtfx': 'median'
}).reset_index()
# Get unique sample counts per model
unique_samples = filtered_df.groupby(['model_family', 'model', 'language', 'dataset_group'])['num_samples'].first().reset_index()
model_samples = unique_samples.groupby(['model_family', 'model'])['num_samples'].sum().reset_index()
model_samples.columns = ['model_family', 'model', 'total_samples']
model_perf.columns = ['model_family', 'model', 'wer_median', 'wer_std', 'count', 'cer_median', 'rtfx_median']
model_perf = model_perf.merge(model_samples, on=['model_family', 'model'], how='left')
model_perf = model_perf.sort_values('wer_median').head(top_n_models)
# Create short model name for display
model_perf['model_short'] = model_perf['model'].apply(
lambda x: x.split('/')[-1] if '/' in x else x
)
# Get colors for each model based on family
model_perf['color'] = model_perf['model_family'].apply(_get_color_for_family)
fig = go.Figure()
fig.add_trace(go.Bar(
y=model_perf['model_short'],
x=model_perf['wer_median'],
orientation='h',
marker=dict(color=model_perf['color']),
text=model_perf['wer_median'].round(2),
textposition='outside',
hovertemplate=(
'<b>%{customdata[0]}</b><br>' +
'<i>Family: %{customdata[1]}</i><br><br>' +
'<b>Median WER:</b> %{x:.3f}<br>' +
'<b>RTFx:</b> %{customdata[2]:.1f}<br>' +
'<b>Evaluations:</b> %{customdata[3]}<br>' +
'<b>Samples:</b> %{customdata[4]:,}<extra></extra>'
),
customdata=model_perf[['model', 'model_family', 'rtfx_median', 'count', 'total_samples']]
))
# Build title with language info (languages is guaranteed to be non-empty here)
lang_list = languages if languages else []
lang_str = ", ".join(lang_list[:3]) + ("..." if len(lang_list) > 3 else "")
title_text = f"Top {min(top_n_models, len(model_perf))} Models for {lang_str}"
else:
# Model family mode (original behavior)
model_perf = filtered_df.groupby('model_family').agg({
'wer': ['median', 'std', 'count'],
'cer': 'median',
'rtfx': 'median'
}).reset_index()
# Get unique sample counts per model family (avoid double-counting across models)
unique_samples = filtered_df.groupby(['model_family', 'language', 'dataset_group'])['num_samples'].first().reset_index()
family_samples = unique_samples.groupby('model_family')['num_samples'].sum().reset_index()
family_samples.columns = ['model_family', 'total_samples']
model_perf.columns = ['model_family', 'wer_median', 'wer_std', 'count', 'cer_median', 'rtfx_median']
model_perf = model_perf.merge(family_samples, on='model_family', how='left')
model_perf = model_perf.sort_values('wer_median')
# Get colors for each model family
model_perf['color'] = model_perf['model_family'].apply(_get_color_for_family)
fig = go.Figure()
fig.add_trace(go.Bar(
y=model_perf['model_family'],
x=model_perf['wer_median'],
orientation='h',
error_x=dict(type='data', array=model_perf['wer_std']),
marker=dict(color=model_perf['color']),
text=model_perf['wer_median'].round(2),
textposition='outside',
hovertemplate=(
'<b>%{y}</b><br><br>' +
'<b>Median WER:</b> %{x:.3f}<br>' +
'<b>Std Dev:</b> %{customdata[0]:.3f}<br>' +
'<b>RTFx:</b> %{customdata[1]:.1f}<br>' +
'<b>Evaluations:</b> %{customdata[2]}<br>' +
'<b>Samples:</b> %{customdata[3]:,}<extra></extra>'
),
customdata=model_perf[['wer_std', 'rtfx_median', 'count', 'total_samples']]
))
title_text = "Model Family Performance Leaderboard"
# Calculate dynamic height based on number of items
num_items = len(model_perf)
height = max(400, min(700, 100 + num_items * 35))
fig.update_layout(
title=title_text,
xaxis_title="Word Error Rate (WER)",
yaxis_title="",
height=height,
autosize=True,
showlegend=False,
template='plotly_white',
margin=dict(l=200, r=30, t=60, b=60)
)
return fig
def create_cer_leaderboard(df: pd.DataFrame, languages: list[str] | None = None, top_n_models: int = 15) -> go.Figure:
"""
Visualization: CER Model Family / Individual Model Performance Leaderboard
- When no language filter: Shows model families (aggregated)
- When language(s) selected: Shows top N individual models
Outliers are removed for cleaner visualization.
Args:
df: DataFrame with evaluation results
languages: List of languages to filter by (None = all languages)
top_n_models: Number of top individual models to show when languages are filtered (default: 15)
"""
# Apply language filter if provided
filtered_df = df.copy()
if languages:
filtered_df = filtered_df[filtered_df['language'].isin(languages)]
# Remove CER outliers for cleaner visualization (similar to WER outlier removal)
if not filtered_df.empty and 'cer' in filtered_df.columns:
Q1 = filtered_df['cer'].quantile(0.25)
Q3 = filtered_df['cer'].quantile(0.75)
IQR = Q3 - Q1
upper_bound = Q3 + 1.5 * IQR
filtered_df = filtered_df[filtered_df['cer'] <= upper_bound]
# Determine mode: individual models if languages selected, otherwise model families
show_individual_models = languages is not None and len(languages) > 0
if show_individual_models:
# Individual model mode: show top N models by median CER
model_perf = filtered_df.groupby(['model_family', 'model']).agg({
'cer': ['median', 'std', 'count'],
'wer': 'median',
'rtfx': 'median'
}).reset_index()
# Get unique sample counts per model
unique_samples = filtered_df.groupby(['model_family', 'model', 'language', 'dataset_group'])['num_samples'].first().reset_index()
model_samples = unique_samples.groupby(['model_family', 'model'])['num_samples'].sum().reset_index()
model_samples.columns = ['model_family', 'model', 'total_samples']
model_perf.columns = ['model_family', 'model', 'cer_median', 'cer_std', 'count', 'wer_median', 'rtfx_median']
model_perf = model_perf.merge(model_samples, on=['model_family', 'model'], how='left')
model_perf = model_perf.sort_values('cer_median').head(top_n_models)
# Create short model name for display
model_perf['model_short'] = model_perf['model'].apply(
lambda x: x.split('/')[-1] if '/' in x else x
)
# Get colors for each model based on family
model_perf['color'] = model_perf['model_family'].apply(_get_color_for_family)
fig = go.Figure()
fig.add_trace(go.Bar(
y=model_perf['model_short'],
x=model_perf['cer_median'],
orientation='h',
marker=dict(color=model_perf['color']),
text=model_perf['cer_median'].round(2),
textposition='outside',
hovertemplate=(
'<b>%{customdata[0]}</b><br>' +
'<i>Family: %{customdata[1]}</i><br><br>' +
'<b>Median CER:</b> %{x:.3f}<br>' +
'<b>WER:</b> %{customdata[2]:.3f}<br>' +
'<b>RTFx:</b> %{customdata[3]:.1f}<br>' +
'<b>Evaluations:</b> %{customdata[4]}<br>' +
'<b>Samples:</b> %{customdata[5]:,}<extra></extra>'
),
customdata=model_perf[['model', 'model_family', 'wer_median', 'rtfx_median', 'count', 'total_samples']]
))
# Build title with language info (languages is guaranteed to be non-empty here)
lang_list = languages if languages else []
lang_str = ", ".join(lang_list[:3]) + ("..." if len(lang_list) > 3 else "")
title_text = f"Top {min(top_n_models, len(model_perf))} Models by CER for {lang_str}"
else:
# Model family mode (original behavior)
model_perf = filtered_df.groupby('model_family').agg({
'cer': ['median', 'std', 'count'],
'wer': 'median',
'rtfx': 'median'
}).reset_index()
# Get unique sample counts per model family (avoid double-counting across models)
unique_samples = filtered_df.groupby(['model_family', 'language', 'dataset_group'])['num_samples'].first().reset_index()
family_samples = unique_samples.groupby('model_family')['num_samples'].sum().reset_index()
family_samples.columns = ['model_family', 'total_samples']
model_perf.columns = ['model_family', 'cer_median', 'cer_std', 'count', 'wer_median', 'rtfx_median']
model_perf = model_perf.merge(family_samples, on='model_family', how='left')
model_perf = model_perf.sort_values('cer_median')
# Get colors for each model family
model_perf['color'] = model_perf['model_family'].apply(_get_color_for_family)
fig = go.Figure()
fig.add_trace(go.Bar(
y=model_perf['model_family'],
x=model_perf['cer_median'],
orientation='h',
error_x=dict(type='data', array=model_perf['cer_std']),
marker=dict(color=model_perf['color']),
text=model_perf['cer_median'].round(2),
textposition='outside',
hovertemplate=(
'<b>%{y}</b><br><br>' +
'<b>Median CER:</b> %{x:.3f}<br>' +
'<b>Std Dev:</b> %{customdata[0]:.3f}<br>' +
'<b>WER:</b> %{customdata[1]:.3f}<br>' +
'<b>RTFx:</b> %{customdata[2]:.1f}<br>' +
'<b>Evaluations:</b> %{customdata[3]}<br>' +
'<b>Samples:</b> %{customdata[4]:,}<extra></extra>'
),
customdata=model_perf[['cer_std', 'wer_median', 'rtfx_median', 'count', 'total_samples']]
))
title_text = "Model Family Performance by CER"
# Calculate dynamic height based on number of items
num_items = len(model_perf)
height = max(400, min(700, 100 + num_items * 35))
fig.update_layout(
title=title_text,
xaxis_title="Character Error Rate (CER)",
yaxis_title="",
height=height,
autosize=True,
showlegend=False,
template='plotly_white',
margin=dict(l=200, r=30, t=60, b=60)
)
return fig
def create_speed_accuracy_scatter(df: pd.DataFrame, view_mode: str = "model_family", languages: list[str] | None = None) -> go.Figure:
"""
Visualization 2: Speed vs Accuracy Tradeoff
Scatter plot showing the relationship between WER and RTFx with quadrants.
Outliers are removed for cleaner visualization.
Args:
df: DataFrame with evaluation results
view_mode: Either "model_family" (bubbles same size per family, color by family) or
"individual_model" (bubble size = model params, color = model family)
languages: List of languages to filter by (None = all languages)
"""
# Apply language filter if provided
if languages:
df = df[df['language'].isin(languages)]
# Remove WER outliers for cleaner visualization
df = _remove_wer_outliers(df)
if view_mode == "individual_model":
# Aggregate by individual model
model_agg = df.groupby(['model_family', 'model']).agg({
'wer': 'median',
'rtfx': 'median',
'cer': 'median',
}).reset_index()
# Get unique sample counts per model
unique_samples = df.groupby(['model_family', 'model', 'language', 'dataset_group'])['num_samples'].first().reset_index()
model_samples = unique_samples.groupby(['model_family', 'model'])['num_samples'].sum().reset_index()
model_samples.columns = ['model_family', 'model', 'num_samples']
model_agg = model_agg.merge(model_samples, on=['model_family', 'model'], how='left')
# Get parameter count for each individual model
model_agg['params'] = model_agg['model'].apply(
lambda m: MODEL_PARAMETER_COUNTS.get(m, 500_000_000) # Default 500M
)
model_agg['params_billions'] = model_agg['params'] / 1_000_000_000
model_agg['params_display'] = model_agg['params'].apply(
lambda x: f"{x/1_000_000_000:.1f}B" if x >= 1_000_000_000 else f"{x/1_000_000:.0f}M"
)
# Create short model name for display
model_agg['model_short'] = model_agg['model'].apply(
lambda x: x.split('/')[-1] if '/' in x else x
)
fig = go.Figure()
# Add scatter traces for each model family
for family in model_agg['model_family'].unique():
family_data = model_agg[model_agg['model_family'] == family]
family_color = _get_color_for_family(family)
fig.add_trace(go.Scatter(
x=family_data['wer'],
y=family_data['rtfx'],
mode='markers',
name=family,
marker=dict(
size=family_data['params'] / family_data['params'].max() * 50 + 10,
sizemode='diameter',
sizemin=8,
color=family_color
),
customdata=family_data[['model_short', 'cer', 'num_samples', 'params_display']].values,
hovertemplate=(
'<b>%{customdata[0]}</b><br><br>' +
'<b>WER:</b> %{x:.3f}<br>' +
'<b>RTFx:</b> %{y:.1f}<br>' +
'<b>CER:</b> %{customdata[1]:.3f}<br>' +
'<b>Samples:</b> %{customdata[2]:,}<br>' +
'<b>Parameters:</b> %{customdata[3]}<extra></extra>'
)
))
title_text = "Speed vs Accuracy Tradeoff by Individual Model"
else:
# Original behavior: aggregate by model family
model_agg = df.groupby('model_family').agg({
'wer': 'median',
'rtfx': 'median',
'cer': 'median',
'model': 'first' # Get a representative model name for parameter lookup
}).reset_index()
# Get unique sample counts per model family
unique_samples = df.groupby(['model_family', 'language', 'dataset_group'])['num_samples'].first().reset_index()
family_samples = unique_samples.groupby('model_family')['num_samples'].sum().reset_index()
family_samples.columns = ['model_family', 'num_samples']
model_agg = model_agg.merge(family_samples, on='model_family', how='left')
# For model family view, use a uniform size (no bubble size variation)
# Use a constant for params to make bubbles the same size
model_agg['params'] = 1_000_000_000 # Use constant 1B for uniform bubble size
model_agg['params_display'] = 'N/A' # Not applicable in family view
fig = go.Figure()
# Add scatter traces for each model family
for family in model_agg['model_family'].unique():
family_data = model_agg[model_agg['model_family'] == family]
family_color = _get_color_for_family(family)
fig.add_trace(go.Scatter(
x=family_data['wer'],
y=family_data['rtfx'],
mode='markers',
name=family,
marker=dict(
size=20,
sizemode='diameter',
color=family_color
),
customdata=family_data[['cer', 'num_samples']].values,
hovertemplate=(
'<b>' + family + '</b><br><br>' +
'<b>WER:</b> %{x:.3f}<br>' +
'<b>RTFx:</b> %{y:.1f}<br>' +
'<b>CER:</b> %{customdata[0]:.3f}<br>' +
'<b>Samples:</b> %{customdata[1]:,}<extra></extra>'
)
))
title_text = "Speed vs Accuracy Tradeoff by Model Family"
# Add quadrant lines at median
median_wer = model_agg['wer'].median()
median_rtfx = model_agg['rtfx'].median()
fig.add_hline(y=median_rtfx, line_dash="dash", line_color="gray", annotation_text="Median RTFx", annotation_position="right")
fig.add_vline(x=median_wer, line_dash="dash", line_color="gray", annotation_text="Median WER", annotation_position="top")
# Add quadrant label - centered in the "Fast & Accurate" quadrant (low WER, high RTFx)
# The ideal quadrant is: x from min to median_wer, y from median_rtfx to max
quadrant_center_x = (model_agg['wer'].min() + median_wer) / 2
quadrant_center_y = (median_rtfx + model_agg['rtfx'].max()) / 2
fig.add_annotation(
x=quadrant_center_x,
y=quadrant_center_y,
text="Fast & Accurate ⭐",
showarrow=False,
font=dict(size=12, color="green", family="Arial Black")
)
fig.update_layout(
title=title_text,
xaxis_title="WER",
yaxis_title="RTFx",
yaxis=dict(rangemode='tozero'), # Ensure y-axis starts at 0 (RTFx can't be negative)
height=550,
autosize=True,
template='plotly_white',
legend=dict(
orientation="h",
yanchor="top",
y=-0.15,
xanchor="center",
x=0.5,
font=dict(size=10)
),
margin=dict(l=60, r=20, t=50, b=120)
)
return fig
def create_wer_cer_correlation(df: pd.DataFrame, languages: list[str] | None = None, top_n_models: int | None = None) -> go.Figure:
"""
Visualization 7: WER vs CER Correlation
Scatter plot showing the relationship between word and character error rates.
Defaults to Swahili if no language is specified.
Outliers are removed for cleaner visualization.
Args:
df: DataFrame with evaluation results
languages: List of languages to filter by (defaults to ["Swahili"] if None)
top_n_models: If specified, only show top N models by WER (0 = show all)
"""
# Default to Swahili if no language filter provided
if not languages:
languages = ["Swahili"]
# Apply language filter
filtered_df = df.copy()
filtered_df = filtered_df[filtered_df['language'].isin(languages)]
# Remove WER outliers for cleaner visualization
filtered_df = _remove_wer_outliers(filtered_df)
# Apply top N models filter if specified
if top_n_models and top_n_models > 0:
model_wer = filtered_df.groupby('model')['wer'].median().sort_values()
top_models = model_wer.head(top_n_models).index.tolist()
filtered_df = filtered_df[filtered_df['model'].isin(top_models)]
fig = go.Figure()
# Add scatter traces for each model family
for family in filtered_df['model_family'].unique():
family_data = filtered_df[filtered_df['model_family'] == family]
family_color = _get_color_for_family(family)
# Normalize size for better visualization
max_samples = filtered_df['num_samples'].max() if not filtered_df.empty else 1
sizes = (family_data['num_samples'] / max_samples * 25 + 5).values if not family_data.empty else [10]
fig.add_trace(go.Scatter(
x=family_data['wer'],
y=family_data['cer'],
mode='markers',
name=family,
marker=dict(
size=sizes,
sizemode='diameter',
sizemin=5,
opacity=0.6,
color=family_color
),
customdata=family_data[['language', 'model', 'dataset_group', 'num_samples']].values,
hovertemplate=(
'<b>%{customdata[0]}</b><br><br>' +
'<b>Model:</b> %{customdata[1]}<br>' +
'<b>Dataset:</b> %{customdata[2]}<br>' +
'<b>WER:</b> %{x:.3f}<br>' +
'<b>CER:</b> %{y:.3f}<br>' +
'<b>Samples:</b> %{customdata[3]:,}<extra></extra>'
)
))
# Add trendline
if len(filtered_df) > 1:
import numpy as np
z = np.polyfit(filtered_df['wer'], filtered_df['cer'], 1)
p = np.poly1d(z)
x_range = np.linspace(filtered_df['wer'].min(), filtered_df['wer'].max(), 100)
fig.add_trace(go.Scatter(
x=x_range,
y=p(x_range),
mode='lines',
name='Trend',
line=dict(color='gray', dash='dash'),
hoverinfo='skip'
))
# Calculate correlation
correlation = filtered_df[['wer', 'cer']].corr().iloc[0, 1] if len(filtered_df) > 1 else 0
# Build title with language info
lang_str = ", ".join(languages[:3]) + ("..." if len(languages) > 3 else "")
title_text = f"WER vs CER Correlation for {lang_str} (r={correlation:.2f})"
fig.update_layout(
title=title_text,
xaxis_title="WER",
yaxis_title="CER",
height=550,
autosize=True,
template='plotly_white',
legend=dict(
orientation="h",
yanchor="top",
y=-0.15,
xanchor="center",
x=0.5,
font=dict(size=10)
),
margin=dict(l=60, r=20, t=50, b=120),
# Zoom X-axis to useful range (0-1.5)
xaxis=dict(range=[0, 1.5]),
yaxis=dict(range=[0, 1.5])
)
return fig
def create_model_consistency(df: pd.DataFrame) -> go.Figure:
"""
Visualization 9: Model Consistency Analysis
Shows coefficient of variation (CV) to measure consistency across languages.
Removes high outliers only using IQR method (keeps best performers).
"""
# Remove only HIGH outliers using IQR method on WER (keep best performers)
Q1 = df['wer'].quantile(0.25)
Q3 = df['wer'].quantile(0.75)
IQR = Q3 - Q1
upper_bound = Q3 + 1.5 * IQR
df_no_outliers = df[df['wer'] <= upper_bound]
model_variance = df_no_outliers.groupby('model_family').agg({
'wer': ['median', 'std', 'count']
}).reset_index()
model_variance.columns = ['model_family', 'wer_median', 'wer_std', 'count']
model_variance['cv'] = (model_variance['wer_std'] / model_variance['wer_median'] * 100)
model_variance = model_variance.sort_values('cv')
# Get colors for each model family
model_variance['color'] = model_variance['model_family'].apply(_get_color_for_family)
fig = go.Figure()
fig.add_trace(go.Bar(
y=model_variance['model_family'],
x=model_variance['cv'],
orientation='h',
marker=dict(
color=model_variance['color']
),
text=model_variance['cv'].round(1),
textposition='outside',
hovertemplate=(
'<b>%{y}</b><br><br>' +
'<b>Coefficient of Variation:</b> %{x:.1f}%<br>' +
'<b>Median WER:</b> %{customdata[0]:.3f}<br>' +
'<b>Std Dev:</b> %{customdata[1]:.3f}<br>' +
'<b>Evaluations:</b> %{customdata[2]}<extra></extra>'
),
customdata=model_variance[['wer_median', 'wer_std', 'count']]
))
fig.update_layout(
title="Model Consistency Ranking (Outliers Removed)",
xaxis_title="Coefficient of Variation (%)",
yaxis_title="Model Family",
height=550,
template='plotly_white',
showlegend=False,
margin=dict(l=200, r=100, t=80, b=80)
)
return fig
|