File size: 36,331 Bytes
0256982 e45f29b 0256982 e45f29b 0256982 e45f29b 0256982 e45f29b 0256982 e45f29b 0256982 e45f29b 0256982 e45f29b 0256982 |
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 |
import streamlit as st
import polars as pl
import altair as alt
from huggingface_hub import hf_hub_download
import traceback
# --- Configuration and Setup ---
st.set_page_config(layout="wide")
alt.data_transformers.enable(
"default"
) # Use default, vegafusion might need server extension
# alt.data_transformers.enable("vegafusion") # Optional: if vegafusion works in your env
st.title("Search Arena V1 Dataset Analysis")
# --- Hugging Face Authentication & Data Loading ---
@st.cache_data(ttl=3600) # Cache data for 1 hour
def load_data():
"""Loads the dataset from Hugging Face Hub."""
try:
repo_id = "lmarena-ai/search-arena-v1-7k"
filename = "data/search-arena-v1-preference-7k.parquet"
# Download the parquet file using the token if provided
# hf_hub_download handles caching locally within the Space's container
local_path = hf_hub_download(
repo_id=repo_id, filename=filename, repo_type="dataset"
)
df = pl.read_parquet(local_path)
df = df.with_columns(pl.col("timestamp").dt.date().alias("date"))
return df
except Exception as e:
st.error(f"Error loading data: {e}")
st.warning(
"Ensure you have added your Hugging Face token (with read access) as a secret named 'HF_TOKEN' in your Space settings."
)
return None
df = load_data()
if df is None:
st.stop() # Stop execution if data loading failed
# --- Data Processing and Visualization ---
st.header("Dataset Overview")
# **Date Histogram**
st.subheader("Data Collection Timeline")
try:
date_counts = df.group_by("date").agg(pl.len().alias("count")).sort("date")
chart_dates = (
alt.Chart(date_counts.to_pandas()) # Convert to Pandas for Altair/Streamlit
.mark_bar()
.encode(
x=alt.X(
"date:T", axis=alt.Axis(labelAngle=0, title="Date")
), # Adjusted angle
y=alt.Y("count:Q", axis=alt.Axis(title="Frequency")),
tooltip=["date:T", "count:Q"],
)
.properties(title="Histogram of Dates")
)
st.altair_chart(chart_dates, use_container_width=True)
except Exception as e:
st.error(f"Error generating date histogram: {e}")
# --- Model Performance Analysis ---
st.header("Model Performance")
# **Calculate Model Statistics**
@st.cache_data(ttl=3600)
def calculate_model_stats(_df):
"""Calculates wins, losses, ties, and win rate for each model."""
unique_models = (
pl.concat([_df["model_a"], _df["model_b"]]).unique().sort().to_list()
)
results = []
for model in unique_models:
# Wins
wins_as_a = _df.filter(
(pl.col("model_a") == model) & (pl.col("winner") == "model_a")
).height
wins_as_b = _df.filter(
(pl.col("model_b") == model) & (pl.col("winner") == "model_b")
).height
total_wins = wins_as_a + wins_as_b
# Losses
losses_as_a = _df.filter(
(pl.col("model_a") == model) & (pl.col("winner") == "model_b")
).height
losses_as_b = _df.filter(
(pl.col("model_b") == model) & (pl.col("winner") == "model_a")
).height
total_losses = losses_as_a + losses_as_b
# Ties (including bothbad)
ties_as_a = _df.filter(
(pl.col("model_a") == model) & (pl.col("winner").str.contains("tie"))
).height
ties_as_b = _df.filter(
(pl.col("model_b") == model) & (pl.col("winner").str.contains("tie"))
).height
total_ties = ties_as_a + ties_as_b
# Total Matches
total_matches = _df.filter(
(pl.col("model_a") == model) | (pl.col("model_b") == model)
).height
# Win Rate
win_rate = (
round(total_wins / total_matches * 100, 2) if total_matches > 0 else 0
)
results.append(
{
"model": model,
"wins": total_wins,
"losses": total_losses,
"ties": total_ties,
"total_matches": total_matches,
"win_rate (%)": win_rate,
}
)
results_df = pl.DataFrame(results).sort("win_rate (%)", descending=True)
return results_df, unique_models
results_df, unique_models = calculate_model_stats(df)
st.subheader("Overall Win Rates")
st.dataframe(results_df.to_pandas(), use_container_width=True) # Display as dataframe
# **Head-to-Head Analysis**
st.subheader("Head-to-Head Matchups (Wins-Losses-Ties-BothBad)")
@st.cache_data(ttl=3600)
def calculate_head_to_head(_df, _unique_models):
"""Calculates head-to-head results."""
head_to_head = []
for model_1 in _unique_models:
row = {"model": model_1}
for model_2 in _unique_models:
if model_1 == model_2:
row[model_2] = "N/A"
continue
matches_ab = _df.filter(
(pl.col("model_a") == model_1) & (pl.col("model_b") == model_2)
)
matches_ba = _df.filter(
(pl.col("model_a") == model_2) & (pl.col("model_b") == model_1)
)
wins_1 = (
matches_ab.filter(pl.col("winner") == "model_a").height
+ matches_ba.filter(pl.col("winner") == "model_b").height
)
wins_2 = (
matches_ab.filter(pl.col("winner") == "model_b").height
+ matches_ba.filter(pl.col("winner") == "model_a").height
)
ties = (
matches_ab.filter(pl.col("winner") == "tie").height
+ matches_ba.filter(pl.col("winner") == "tie").height
)
bothbad = (
matches_ab.filter(pl.col("winner") == "tie (bothbad)").height
+ matches_ba.filter(pl.col("winner") == "tie (bothbad)").height
)
total = wins_1 + wins_2 + ties + bothbad
row[model_2] = (
f"{wins_1}-{wins_2}-{ties}-{bothbad}" if total > 0 else "0-0-0-0"
)
head_to_head.append(row)
return pl.DataFrame(head_to_head)
head_to_head_df = calculate_head_to_head(df, unique_models)
st.dataframe(head_to_head_df.to_pandas(), use_container_width=True)
# **Heatmaps**
st.subheader("Head-to-Head Heatmaps")
@st.cache_data(ttl=3600)
def prepare_heatmap_data(_head_to_head_df):
"""Prepares data for heatmaps."""
melted_df = _head_to_head_df.unpivot(index=["model"], variable_name="opponent")
parsed_data = []
for row in melted_df.iter_rows(named=True):
model, opponent, value = row["model"], row["opponent"], row["value"]
if value != "N/A":
try:
parts = value.split("-")
wins, losses, ties, bothbad = map(int, parts)
parsed_data.extend(
[
{
"model": model,
"opponent": opponent,
"metric": "wins",
"value": wins,
},
{
"model": model,
"opponent": opponent,
"metric": "losses",
"value": losses,
},
{
"model": model,
"opponent": opponent,
"metric": "ties",
"value": ties,
},
{
"model": model,
"opponent": opponent,
"metric": "bothbad",
"value": bothbad,
},
]
)
except (ValueError, IndexError):
st.warning(
f"Could not parse head-to-head value: '{value}' for {model} vs {opponent}"
)
continue # Skip malformed entries
return pl.DataFrame(parsed_data)
metrics_df = prepare_heatmap_data(head_to_head_df)
def create_heatmap(data_pd, metric, color_scheme):
"""Helper function to create an Altair heatmap."""
if data_pd.empty:
return alt.Chart(pd.DataFrame({"x": [], "y": [], "value": []})).mark_text(
text=f"No data for {metric}"
)
median_value = data_pd["value"].median()
heatmap = (
alt.Chart(data_pd)
.mark_rect()
.encode(
x=alt.X(
"opponent:N",
title="Opponent",
sort=unique_models,
axis=alt.Axis(labelLimit=0, labelAngle=90),
), # Ensure consistent sorting
y=alt.Y(
"model:N", title="Model", sort=unique_models
), # Ensure consistent sorting
color=alt.Color(
"value:Q",
scale=alt.Scale(scheme=color_scheme),
title=f"{metric.capitalize()}",
),
tooltip=["model", "opponent", "value"],
)
.properties(
title=f"{metric.capitalize()}", width=alt.Step(40), height=alt.Step(40)
) # Adjust step for size
)
text = heatmap.mark_text(baseline="middle").encode(
text="value:Q",
color=alt.condition(
f"datum.value > {median_value}", alt.value("white"), alt.value("black")
),
tooltip=["model", "opponent", "value"], # Ensure tooltip is on text too
)
return heatmap + text
# Filter and convert data for each heatmap within Streamlit rendering flow
col1, col2 = st.columns(2)
with col1:
try:
wins_pd = metrics_df.filter(pl.col("metric") == "wins").to_pandas()
if not wins_pd.empty:
wins_heatmap = create_heatmap(wins_pd, "wins", "greens")
st.altair_chart(wins_heatmap, use_container_width=True)
else:
st.write("No 'wins' data for heatmap.")
except Exception as e:
st.error(f"Error creating wins heatmap: {e}")
try:
ties_pd = metrics_df.filter(pl.col("metric") == "ties").to_pandas()
if not ties_pd.empty:
ties_heatmap = create_heatmap(ties_pd, "ties", "purples")
st.altair_chart(ties_heatmap, use_container_width=True)
else:
st.write("No 'ties' data for heatmap.")
except Exception as e:
st.error(f"Error creating ties heatmap: {e}")
with col2:
try:
losses_pd = metrics_df.filter(pl.col("metric") == "losses").to_pandas()
if not losses_pd.empty:
losses_heatmap = create_heatmap(losses_pd, "losses", "oranges")
st.altair_chart(losses_heatmap, use_container_width=True)
else:
st.write("No 'losses' data for heatmap.")
except Exception as e:
st.error(f"Error creating losses heatmap: {e}")
try:
bothbad_pd = metrics_df.filter(pl.col("metric") == "bothbad").to_pandas()
if not bothbad_pd.empty:
bothbad_heatmap = create_heatmap(bothbad_pd, "bothbad", "reds")
st.altair_chart(bothbad_heatmap, use_container_width=True)
else:
st.write("No 'bothbad' data for heatmap.")
except Exception as e:
st.error(f"Error creating bothbad heatmap: {e}")
# --- Detailed Outcome Analysis ---
st.header("Detailed Outcome Analysis")
# **Reshape Data for Detailed Analysis**
@st.cache_data(ttl=3600)
def reshape_data_for_analysis(_df):
"""Reshapes data for outcome analysis per model."""
# Helper to safely extract struct fields, returning None if field doesn't exist or is null
def safe_struct_field(col_name, field_name):
return (
pl.when(pl.col(col_name).struct.field(field_name).is_not_null())
.then(pl.col(col_name).struct.field(field_name))
.otherwise(None)
) # Or pl.lit(0) or appropriate default
df_model_a = _df.select(
[
pl.col("model_a").alias("model"),
pl.when(pl.col("winner") == "model_a")
.then(pl.lit("win"))
.when(pl.col("winner") == "tie")
.then(pl.lit("tie"))
.when(pl.col("winner") == "tie (bothbad)")
.then(pl.lit("tie (bothbad)"))
.otherwise(pl.lit("loss"))
.alias("outcome"),
"language",
"turn",
"date",
safe_struct_field("conv_metadata", "response_length_a").alias(
"response_length"
),
safe_struct_field("conv_metadata", "num_citations_a").alias(
"num_citations"
),
]
)
df_model_b = _df.select(
[
pl.col("model_b").alias("model"),
pl.when(pl.col("winner") == "model_b")
.then(pl.lit("win"))
.when(pl.col("winner") == "tie")
.then(pl.lit("tie"))
.when(pl.col("winner") == "tie (bothbad)")
.then(pl.lit("tie (bothbad)"))
.otherwise(pl.lit("loss"))
.alias("outcome"),
"language",
"turn",
"date",
safe_struct_field("conv_metadata", "response_length_b").alias(
"response_length"
),
safe_struct_field("conv_metadata", "num_citations_b").alias(
"num_citations"
),
]
)
# Drop rows where essential analysis columns are null if necessary
df_models = pl.concat([df_model_a, df_model_b], how="vertical")
df_models = df_models.fill_null(
0
) # Fill nulls after concat, maybe with 0 or a specific strategy
return df_models
df_models = reshape_data_for_analysis(df)
# **Calculate Rates**
@st.cache_data(ttl=3600)
def calculate_rates(_df_models):
"""Calculates various outcome rates per model."""
df_rates = (
_df_models.group_by("model")
.agg(
wins=pl.col("outcome").eq("win").sum(),
losses=pl.col("outcome").eq("loss").sum(),
ties=pl.col("outcome").eq("tie").sum(),
tie_bothbad=pl.col("outcome").eq("tie (bothbad)").sum(),
)
.with_columns(
total=pl.sum_horizontal(["wins", "losses", "ties", "tie_bothbad"]),
)
.filter(pl.col("total") > 0) # Avoid division by zero
.with_columns(
win_rate=pl.col("wins") / pl.col("total"),
loss_rate=pl.col("losses") / pl.col("total"),
tie_rate=pl.col("ties") / pl.col("total"),
tie_bothbad_rate=pl.col("tie_bothbad") / pl.col("total"),
weighted_rate=(
(pl.col("wins") + 0.5 * pl.col("ties") + 0.25 * pl.col("tie_bothbad"))
/ pl.col("total")
),
)
)
return df_rates
df_rates = calculate_rates(df_models)
# **Outcome Distribution Chart**
st.subheader("Outcome Distribution by Model")
try:
outcome_order = ["win", "loss", "tie", "tie (bothbad)"]
winner_bar = (
alt.Chart(df_models.to_pandas())
.mark_bar()
.encode(
x=alt.X(
"model:N",
title="Model",
sort=unique_models,
axis=alt.Axis(labelLimit=0),
),
y=alt.Y("count():Q", title="Count"),
color=alt.Color(
"outcome:N",
title="Outcome",
scale=alt.Scale(
domain=outcome_order,
range=["#1f77b4", "#d62728", "#2ca02c", "#9467bd"],
),
sort=outcome_order,
), # Explicit colors/order
tooltip=["model", "outcome", "count()"],
order=alt.Order(
"color_outcome_sort_index:Q"
), # Ensure stack order matches legend
)
.transform_calculate(
# Create a field for sorting based on the domain order
color_outcome_sort_index=f"{{'win': 0, 'loss': 1, 'tie': 2, 'tie (bothbad)': 3}}[datum.outcome]"
)
)
st.altair_chart(winner_bar, use_container_width=True)
except Exception as e:
st.error(f"Error generating outcome distribution chart: {e}")
# **Rates Chart**
st.subheader("Outcome Rates by Model")
try:
df_rates_long = (
df_rates.select(
["model", "win_rate", "loss_rate", "tie_rate", "tie_bothbad_rate"]
)
.unpivot(
index=["model"],
variable_name="rate_type",
value_name="rate_value",
)
.with_columns(
# Clean up rate type names for display
pl.col("rate_type").str.replace("_rate", "").str.to_titlecase()
)
)
rate_order = ["Win", "Loss", "Tie", "Tie Bothbad"] # Order for stacking and legend
stacked_bar = (
alt.Chart(df_rates_long.to_pandas())
.mark_bar()
.encode(
x=alt.X(
"model:N",
title="Model",
sort=unique_models,
axis=alt.Axis(labelLimit=0),
),
y=alt.Y(
"rate_value:Q",
title="Rate",
stack="normalize",
axis=alt.Axis(format="%"),
), # Normalize stack
color=alt.Color("rate_type:N", title="Rate Type", sort=rate_order),
order=alt.Order(
"color_rate_type_sort_index:Q"
), # Use calculation for stack order
tooltip=["model", "rate_type", alt.Tooltip("rate_value:Q", format=".1%")],
)
.transform_calculate(
# Create a field for sorting based on the domain order
color_rate_type_sort_index=f"{{'Win': 0, 'Loss': 1, 'Tie': 2, 'Tie Bothbad': 3}}[datum.rate_type]"
)
)
weighted_line = (
alt.Chart(df_rates.to_pandas())
.mark_line(point=True, color="orange", strokeDash=[5, 5]) # Dashed line
.encode(
x=alt.X("model:N", title="Model", sort=unique_models),
y=alt.Y(
"weighted_rate:Q", title="Weighted Rate", axis=alt.Axis(format=".1%")
),
tooltip=[
"model",
alt.Tooltip("weighted_rate:Q", title="Weighted Rate", format=".1%"),
],
)
)
rates_chart = (
(stacked_bar + weighted_line)
.properties(title="Stacked Outcome Rates by Model (Weighted Rate Overlay)")
.resolve_scale(y="independent")
) # Independent Y-axis for line vs bars
st.altair_chart(rates_chart, use_container_width=True)
except Exception as e:
st.error(f"Error generating rates chart: {e}")
# --- Multilingual Performance ---
st.header("Multilingual Performance")
# **Language Frequency**
st.subheader("Language Distribution")
try:
# Calculate language frequency from df_models which has one row per model appearance
lang_freq_df = (
df_models["language"].value_counts().rename({"count": "total_samples"})
)
language_freq_chart = (
alt.Chart(lang_freq_df.to_pandas())
.mark_bar()
.encode(
x=alt.X("language:N", sort="-y", title="Language"),
y=alt.Y("total_samples:Q", title="Number of Comparisons"),
tooltip=["language", "total_samples"],
)
)
st.altair_chart(language_freq_chart, use_container_width=True)
except Exception as e:
st.error(f"Error generating language frequency chart: {e}")
# **Win Rate by Language Heatmap**
st.subheader("Win Rate by Model and Language")
try:
win_rate_lang_df = (
df_models.group_by(["model", "language"])
.agg(wins=(pl.col("outcome") == "win").sum(), total=pl.len())
.filter(pl.col("total") > 0) # Avoid division by zero
.with_columns(win_rate=pl.col("wins") / pl.col("total"))
)
win_rate_language_heatmap = (
alt.Chart(win_rate_lang_df.to_pandas())
.mark_rect()
.encode(
x=alt.X(
"model:N",
title="Model",
sort=unique_models,
axis=alt.Axis(labelLimit=0),
),
y=alt.Y("language:N", title="Language"),
color=alt.Color(
"win_rate:Q",
title="Win Rate",
scale=alt.Scale(scheme="blues"),
legend=alt.Legend(format=".0%"),
),
tooltip=[
"model",
"language",
alt.Tooltip("win_rate:Q", format=".1%"),
"total",
],
)
)
st.altair_chart(win_rate_language_heatmap, use_container_width=True)
except Exception as e:
st.error(f"Error generating win rate by language heatmap: {e}")
# **Outcome Distribution by Language and Model (Faceted & Wrapped)**
st.subheader("Outcome Distribution by Language and Model")
try:
# --- Data Prep for Annotation (as per working example) ---
@st.cache_data(ttl=3600)
def prepare_data_for_language_facet(_df_models):
df_language_totals = _df_models.group_by("language").agg(total_samples=pl.len())
df_models_with_totals = _df_models.join(
df_language_totals, on="language", how="left"
)
return df_models_with_totals
df_models_for_facet = prepare_data_for_language_facet(df_models)
# Convert to Pandas for Altair
df_models_pd = df_models_for_facet.to_pandas()
# --- Base Bar Chart (using the full dataset) ---
bar_chart = (
alt.Chart(df_models_pd) # Use the common DataFrame
.mark_bar()
.encode(
x=alt.X(
"model:N",
title="Model",
axis=alt.Axis(labels=True, labelLimit=0),
), # Show labels, rotated
# --- Choose one Y encoding ---
# For absolute counts (like example):
y=alt.Y("count():Q", title="Count", stack="zero"),
# For normalized bars:
# y=alt.Y("count():Q", title="%", stack="normalize", axis=alt.Axis(format="%")),
# -----------------------------
color=alt.Color(
"outcome:N",
title="Outcome",
sort=outcome_order, # Use the existing outcome_order list
scale=alt.Scale(
domain=outcome_order,
range=[
"green",
"orange",
"lightblue",
"red",
],
),
),
tooltip=[
"model",
"language",
"outcome",
"count()",
alt.Tooltip("total_samples:Q", title="Total Samples in Language"),
],
order=alt.Order( # Important for consistent stacking, esp. if normalized
"color_outcome_sort_index:Q"
),
)
.transform_calculate(
# Field for sorting stack order
color_outcome_sort_index=f"{{'win': 0, 'loss': 1, 'tie': 2, 'tie (bothbad)': 3}}[datum.outcome]"
)
)
# --- Text Annotation (using the full dataset with transform_aggregate) ---
text_chart = (
alt.Chart(df_models_pd) # Use the *same* common DataFrame
.mark_text(
align="left", baseline="middle", dx=5, dy=-5, color="black", fontSize=10
) # Adjusted position/appearance
.encode(
# Position text relative to the top-left corner of the facet using values
x=alt.value(5), # Small offset from the left edge
y=alt.value(15), # Small offset from the top edge
text=alt.Text("total_samples:Q", format=",d"),
color=alt.value("black"), # Explicit text color
# Tooltip for the text itself (optional)
tooltip=[alt.Tooltip("total_samples:Q", title="Total Samples in Language")],
)
.transform_aggregate(
# Aggregate within Altair to get max total_samples per language
# Note: total_samples is constant per language, so max() is just a way to get it once
total_samples="max(total_samples)",
groupby=["language"], # Group by the faceting variable
)
)
# --- Layer the charts THEN Facet (Correct Approach) ---
language_outcome_chart = (
# Layer first using '+'
(bar_chart + text_chart)
.facet(
# Then facet the layered chart
facet=alt.Facet(
"language:N",
title="Language",
header=alt.Header(titleOrient="top", labelOrient="top"),
),
columns=5, # Wrap into 3 columns
)
.resolve_scale(
y="independent" # Resolve Y scale because text has different positioning logic
)
)
st.altair_chart(language_outcome_chart, use_container_width=True)
except Exception as e:
import traceback # Make sure traceback is imported at the top of your script
st.error(
f"Error generating faceted outcome distribution by language: {e}\n{traceback.format_exc()}"
)
# --- Performance Over Turns and Time ---
st.header("Performance Dynamics")
# **Win Rate by Turn Heatmap**
st.subheader("Win Rate by Turn")
try:
win_rate_turn_df = (
df_models.group_by(["model", "turn"])
.agg(wins=(pl.col("outcome") == "win").sum(), total=pl.len())
.filter(pl.col("total") > 0)
.with_columns(win_rate=pl.col("wins") / pl.col("total"))
)
win_rate_turn_heatmap = (
alt.Chart(win_rate_turn_df.to_pandas())
.mark_rect()
.encode(
x=alt.X("turn:O", title="Turn"), # Treat turn as ordinal
y=alt.Y(
"model:N",
title="Model",
sort=unique_models,
axis=alt.Axis(labelLimit=0),
),
color=alt.Color(
"win_rate:Q",
title="Win Rate",
scale=alt.Scale(scheme="blues"),
legend=alt.Legend(format=".0%"),
),
tooltip=[
"model",
"turn",
alt.Tooltip("win_rate:Q", format=".1%"),
"total",
],
)
.properties(title="Win Rate by Model and Turn")
)
st.altair_chart(win_rate_turn_heatmap, use_container_width=True)
except Exception as e:
st.error(f"Error generating win rate by turn heatmap: {e}")
# **Wins Over Time Line Chart**
st.subheader("Wins Over Time")
try:
wins_time_df = (
df_models.filter(pl.col("outcome") == "win")
.group_by(["date", "model"])
.agg(win_count=pl.len())
.sort("date")
)
time_line = (
alt.Chart(wins_time_df.to_pandas())
.mark_line(point=True)
.encode(
x=alt.X("date:T", title="Date"),
y=alt.Y("win_count:Q", title="Daily Win Count"),
color=alt.Color("model:N", title="Model"),
tooltip=["model", "date", "win_count"],
)
.properties(title="Wins by Model Over Time")
)
st.altair_chart(time_line, use_container_width=True)
except Exception as e:
st.error(f"Error generating wins over time chart: {e}")
# --- Response Characteristics ---
st.header("Response Characteristics vs. Outcome")
# **Response Length Boxplot (Faceted, No Outliers, Adjusted Y-Axis, Detailed Tooltip)**
st.subheader(
"Response Length Distribution by Outcome (Box Plot - Adjusted Scale, No Outliers)"
)
try:
# Filter using .and_() for logical AND
response_length_df = df_models.filter(
pl.col("response_length").is_not_null().and_(pl.col("response_length") >= 0)
)
# Convert to Pandas for Altair and percentile calculation
response_length_pd = response_length_df.to_pandas()
# --- Calculate a reasonable upper limit for the Y-axis ---
valid_lengths = response_length_pd["response_length"].dropna()
valid_lengths = valid_lengths[valid_lengths >= 0]
if not valid_lengths.empty:
upper_limit = valid_lengths.quantile(0.99)
upper_limit = max(upper_limit, 100) # Example minimum range
else:
upper_limit = 1000 # Default if no valid data
st.write(
f"(Note: Y-axis for Response Length capped at ~{int(upper_limit)} [99th percentile] to improve visibility)"
)
response_facet_boxplot = (
alt.Chart(response_length_pd)
.mark_boxplot(
extent="min-max", # Keep whiskers extending to min/max within 1.5*IQR
outliers=False, # Keep outliers hidden
)
.encode(
x=alt.X(
"model:N",
title="Model",
sort=unique_models,
axis=alt.Axis(labelLimit=0),
), # Keep axis formatting
y=alt.Y(
"response_length:Q",
title="Response Length",
scale=alt.Scale(domain=[0, upper_limit], clamp=True), # Keep clamping
),
color=alt.Color(
"model:N", title="Model", legend=None
), # Color by model for clarity within facet
)
.facet(
# Keep the same faceting by outcome
column=alt.Column("outcome:N", title="Outcome", sort=outcome_order)
)
.properties(
title="Response Length Distribution by Model per Outcome (Box Plot, Adjusted Scale, No Outliers)"
)
)
st.altair_chart(response_facet_boxplot, use_container_width=True)
except Exception as e:
import traceback # Ensure traceback is imported
st.error(f"Error generating response length boxplot: {e}\n{traceback.format_exc()}")
# --- Citations Scatter Plot (Faceted) ---
st.subheader("Number of Citations by Outcome")
try:
# Corrected filter using .and_() for logical AND
citation_df = df_models.filter(
pl.col("num_citations").is_not_null().and_(pl.col("num_citations") >= 0)
)
# Aggregate counts for bubble size
citation_agg = citation_df.group_by(["model", "outcome", "num_citations"]).agg(
count=pl.len()
)
# Convert to Pandas if needed
citation_agg_pd = citation_agg.to_pandas()
citation_facet = (
alt.Chart(citation_agg_pd)
.mark_circle()
.encode(
x=alt.X(
"model:N",
title="Model",
sort=unique_models,
axis=alt.Axis(labelLimit=0),
), # Keep axis formatting
y=alt.Y(
"num_citations:Q",
title="Number of Citations",
axis=alt.Axis(tickMinStep=1),
), # Ensure integer ticks
# --- Modify the legend within alt.Size ---
size=alt.Size(
"count:Q",
title="Number of Responses",
legend=alt.Legend(
symbolFillColor="lightblue", # Set legend symbol fill color to white
symbolStrokeColor="lightblue", # Optional: add a subtle border if needed
),
),
# ------------------------------------------
color=alt.Color(
"model:N", title="Model", legend=None
), # Keep model legend hidden
tooltip=["model", "outcome", "num_citations", "count"],
)
.facet(column=alt.Column("outcome:N", title="Outcome", sort=outcome_order))
.properties(title="Citations by Model per Outcome")
)
st.altair_chart(citation_facet, use_container_width=True)
except Exception as e:
import traceback # Ensure traceback is imported
st.error(f"Error generating citation facet plot: {e}\n{traceback.format_exc()}")
# --- Language Leaderboard ---
st.header("Top/Worst Models per Language")
@st.cache_data(ttl=3600)
def calculate_language_ranks(_df_models):
"""Calculates top/worst models based on win rate per language."""
df_win_rates = (
_df_models.group_by(["language", "model"])
.agg(wins=pl.col("outcome").eq("win").sum(), total=pl.len())
.filter(pl.col("total") > 5) # Require minimum samples for ranking
.with_columns(win_rate=pl.col("wins") / pl.col("total"))
)
if df_win_rates.is_empty():
return pl.DataFrame() # Return empty if no language meets criteria
df_ranked = df_win_rates.with_columns(
rank_top=pl.col("win_rate")
.rank(method="min", descending=True)
.over("language"),
rank_worst=pl.col("win_rate")
.rank(method="min", descending=False)
.over("language"),
)
df_top_1 = (
df_ranked.filter(pl.col("rank_top") == 1)
.group_by("language")
.agg(
top_model_1=pl.col("model").first(),
win_rate_1=pl.col("win_rate").first(),
total_samples=pl.col("total").sum(), # Sum totals for the language
)
)
df_top_2 = (
df_ranked.filter(pl.col("rank_top") == 2)
.group_by("language")
.agg(top_model_2=pl.col("model").first(), win_rate_2=pl.col("win_rate").first())
)
df_worst_1 = (
df_ranked.filter(pl.col("rank_worst") == 1)
.group_by("language")
.agg(
worst_model_1=pl.col("model").first(),
worst_win_rate_1=pl.col("win_rate").first(),
)
)
df_worst_2 = (
df_ranked.filter(pl.col("rank_worst") == 2)
.group_by("language")
.agg(
worst_model_2=pl.col("model").first(),
worst_win_rate_2=pl.col("win_rate").first(),
)
)
# Combine using outer joins to handle cases where ranks don't exist (e.g., < 4 models)
df_table = (
df_top_1.join(df_top_2, on="language", how="left")
.join(df_worst_1, on="language", how="left")
.join(df_worst_2, on="language", how="left")
.select(
[
"language",
"total_samples",
"top_model_1",
pl.col("win_rate_1").round(3).alias("WR #1"),
"top_model_2",
pl.col("win_rate_2").round(3).alias("WR #2"),
"worst_model_1",
pl.col("worst_win_rate_1").round(3).alias("WR Worst #1"),
"worst_model_2",
pl.col("worst_win_rate_2").round(3).alias("WR Worst #2"),
]
)
.sort("language")
.fill_null("N/A")
) # Fill missing ranks with N/A
return df_table
df_language_table = calculate_language_ranks(df_models)
if not df_language_table.is_empty():
st.subheader("Top & Bottom 2 Models by Win Rate per Language (Min 5 Comparisons)")
st.dataframe(df_language_table.to_pandas(), use_container_width=True)
else:
st.subheader("Top & Bottom 2 Models by Win Rate per Language")
st.write(
"Insufficient data (fewer than 5 comparisons) for one or more languages to generate rankings."
)
# **Model Ranking Counts**
if not df_language_table.is_empty():
st.subheader("How Often Models Rank Top/Worst Across Languages")
try:
df_top_1_counts = (
df_language_table["top_model_1"]
.value_counts()
.rename({"top_model_1": "model", "count": "Rank 1 Count"})
.sort("Rank 1 Count", descending=True)
)
df_worst_1_counts = (
df_language_table["worst_model_1"]
.value_counts()
.rename({"worst_model_1": "model", "count": "Worst Rank Count"})
.sort("Worst Rank Count", descending=True)
)
col1, col2 = st.columns(2)
with col1:
st.write("**Times Ranked #1**")
st.dataframe(df_top_1_counts.to_pandas(), use_container_width=True)
with col2:
st.write("**Times Ranked Worst**")
st.dataframe(df_worst_1_counts.to_pandas(), use_container_width=True)
except Exception as e:
st.error(f"Error generating model ranking counts: {e}")
# --- Footer ---
st.markdown("---")
st.markdown(
"Analysis based on the `lmarena-ai/search-arena-v1-7k` dataset on Hugging Face."
)
|