Spaces:
Runtime error
Runtime error
Commit
·
33739cc
1
Parent(s):
1d6f829
[REMOVE] Model symbols
Browse files- app.py +94 -92
- src/display/utils.py +1 -1
- src/leaderboard/read_evals.py +1 -1
- src/populate.py +1 -1
app.py
CHANGED
|
@@ -239,13 +239,14 @@ def update_table(
|
|
| 239 |
hidden_df: pd.DataFrame,
|
| 240 |
columns: list,
|
| 241 |
query: str = "",
|
| 242 |
-
type_query: list = None,
|
| 243 |
domain_specific_query: list = None,
|
| 244 |
size_query: list = None,
|
| 245 |
precision_query: str = None,
|
| 246 |
show_deleted: bool = False,
|
| 247 |
):
|
| 248 |
# breakpoint()
|
|
|
|
| 249 |
filtered_df = filter_models(hidden_df, type_query, domain_specific_query, size_query, precision_query, show_deleted)
|
| 250 |
# breakpoint()
|
| 251 |
filtered_df = filter_queries(query, filtered_df)
|
|
@@ -338,7 +339,7 @@ with demo:
|
|
| 338 |
print("hello")
|
| 339 |
gr.HTML(LOGO)
|
| 340 |
gr.Markdown(INTRODUCTION_TEXT, elem_classes="markdown-text")
|
| 341 |
-
|
| 342 |
with gr.Tabs(elem_classes="tab-buttons") as tabs:
|
| 343 |
with gr.TabItem("🏅 Open Ended Evaluation", elem_id="llm-benchmark-tab-table", id=1):
|
| 344 |
with gr.Tabs(elem_classes="tab-buttons6") as language_tabs:
|
|
@@ -383,13 +384,14 @@ with demo:
|
|
| 383 |
interactive=True,
|
| 384 |
)
|
| 385 |
with gr.Column(min_width=320):
|
| 386 |
-
filter_columns_type = gr.CheckboxGroup(
|
| 387 |
-
|
| 388 |
-
|
| 389 |
-
|
| 390 |
-
|
| 391 |
-
|
| 392 |
-
)
|
|
|
|
| 393 |
filter_domain_specific = gr.CheckboxGroup(
|
| 394 |
label="Domain Specificity",
|
| 395 |
choices=["🏥 Clinical models", "Generic models"],
|
|
@@ -429,7 +431,7 @@ with demo:
|
|
| 429 |
hidden_leaderboard_table_for_search,
|
| 430 |
shown_columns,
|
| 431 |
search_bar,
|
| 432 |
-
filter_columns_type,
|
| 433 |
filter_domain_specific,
|
| 434 |
filter_columns_size
|
| 435 |
],
|
|
@@ -438,7 +440,7 @@ with demo:
|
|
| 438 |
|
| 439 |
for selector in [
|
| 440 |
shown_columns,
|
| 441 |
-
filter_columns_type,
|
| 442 |
filter_domain_specific,
|
| 443 |
filter_columns_size,
|
| 444 |
]:
|
|
@@ -448,7 +450,7 @@ with demo:
|
|
| 448 |
hidden_leaderboard_table_for_search,
|
| 449 |
shown_columns,
|
| 450 |
search_bar,
|
| 451 |
-
filter_columns_type,
|
| 452 |
filter_domain_specific,
|
| 453 |
filter_columns_size
|
| 454 |
],
|
|
@@ -490,13 +492,13 @@ with demo:
|
|
| 490 |
# )
|
| 491 |
with gr.Column(min_width=320):
|
| 492 |
# with gr.Box(elem_id="box-filter"):
|
| 493 |
-
filter_columns_type = gr.CheckboxGroup(
|
| 494 |
-
|
| 495 |
-
|
| 496 |
-
|
| 497 |
-
|
| 498 |
-
|
| 499 |
-
)
|
| 500 |
# filter_columns_architecture = gr.CheckboxGroup(
|
| 501 |
# label="Architecture Types",
|
| 502 |
# choices=[i.value.name for i in ModelArch],
|
|
@@ -545,7 +547,7 @@ with demo:
|
|
| 545 |
hidden_leaderboard_table_for_search,
|
| 546 |
shown_columns,
|
| 547 |
search_bar,
|
| 548 |
-
filter_columns_type,
|
| 549 |
filter_domain_specific,
|
| 550 |
filter_columns_size
|
| 551 |
# filter_columns_architecture
|
|
@@ -554,7 +556,7 @@ with demo:
|
|
| 554 |
)
|
| 555 |
for selector in [
|
| 556 |
shown_columns,
|
| 557 |
-
filter_columns_type,
|
| 558 |
filter_domain_specific,
|
| 559 |
filter_columns_size,
|
| 560 |
# deleted_models_visibility,
|
|
@@ -565,7 +567,7 @@ with demo:
|
|
| 565 |
hidden_leaderboard_table_for_search,
|
| 566 |
shown_columns,
|
| 567 |
search_bar,
|
| 568 |
-
filter_columns_type,
|
| 569 |
filter_domain_specific,
|
| 570 |
filter_columns_size
|
| 571 |
],
|
|
@@ -610,13 +612,13 @@ with demo:
|
|
| 610 |
# )
|
| 611 |
with gr.Column(min_width=320):
|
| 612 |
# with gr.Box(elem_id="box-filter"):
|
| 613 |
-
filter_columns_type = gr.CheckboxGroup(
|
| 614 |
-
|
| 615 |
-
|
| 616 |
-
|
| 617 |
-
|
| 618 |
-
|
| 619 |
-
)
|
| 620 |
# filter_columns_architecture = gr.CheckboxGroup(
|
| 621 |
# label="Architecture Types",
|
| 622 |
# choices=[i.value.name for i in ModelArch],
|
|
@@ -665,7 +667,7 @@ with demo:
|
|
| 665 |
hidden_leaderboard_table_for_search,
|
| 666 |
shown_columns,
|
| 667 |
search_bar,
|
| 668 |
-
filter_columns_type,
|
| 669 |
filter_domain_specific,
|
| 670 |
filter_columns_size
|
| 671 |
# filter_columns_architecture
|
|
@@ -674,7 +676,7 @@ with demo:
|
|
| 674 |
)
|
| 675 |
for selector in [
|
| 676 |
shown_columns,
|
| 677 |
-
filter_columns_type,
|
| 678 |
filter_domain_specific,
|
| 679 |
filter_columns_size,
|
| 680 |
# deleted_models_visibility,
|
|
@@ -685,7 +687,7 @@ with demo:
|
|
| 685 |
hidden_leaderboard_table_for_search,
|
| 686 |
shown_columns,
|
| 687 |
search_bar,
|
| 688 |
-
filter_columns_type,
|
| 689 |
filter_domain_specific,
|
| 690 |
filter_columns_size
|
| 691 |
],
|
|
@@ -719,13 +721,13 @@ with demo:
|
|
| 719 |
# )
|
| 720 |
with gr.Column(min_width=320):
|
| 721 |
# with gr.Box(elem_id="box-filter"):
|
| 722 |
-
filter_columns_type = gr.CheckboxGroup(
|
| 723 |
-
|
| 724 |
-
|
| 725 |
-
|
| 726 |
-
|
| 727 |
-
|
| 728 |
-
)
|
| 729 |
# filter_columns_architecture = gr.CheckboxGroup(
|
| 730 |
# label="Architecture Types",
|
| 731 |
# choices=[i.value.name for i in ModelArch],
|
|
@@ -774,7 +776,7 @@ with demo:
|
|
| 774 |
hidden_leaderboard_table_for_search,
|
| 775 |
shown_columns,
|
| 776 |
search_bar,
|
| 777 |
-
filter_columns_type,
|
| 778 |
filter_domain_specific,
|
| 779 |
filter_columns_size
|
| 780 |
# filter_columns_architecture
|
|
@@ -783,7 +785,7 @@ with demo:
|
|
| 783 |
)
|
| 784 |
for selector in [
|
| 785 |
shown_columns,
|
| 786 |
-
filter_columns_type,
|
| 787 |
filter_domain_specific,
|
| 788 |
filter_columns_size,
|
| 789 |
# deleted_models_visibility,
|
|
@@ -794,7 +796,7 @@ with demo:
|
|
| 794 |
hidden_leaderboard_table_for_search,
|
| 795 |
shown_columns,
|
| 796 |
search_bar,
|
| 797 |
-
filter_columns_type,
|
| 798 |
filter_domain_specific,
|
| 799 |
filter_columns_size
|
| 800 |
],
|
|
@@ -841,13 +843,13 @@ with demo:
|
|
| 841 |
# )
|
| 842 |
with gr.Column(min_width=320):
|
| 843 |
# with gr.Box(elem_id="box-filter"):
|
| 844 |
-
filter_columns_type = gr.CheckboxGroup(
|
| 845 |
-
|
| 846 |
-
|
| 847 |
-
|
| 848 |
-
|
| 849 |
-
|
| 850 |
-
)
|
| 851 |
# filter_columns_architecture = gr.CheckboxGroup(
|
| 852 |
# label="Architecture Types",
|
| 853 |
# choices=[i.value.name for i in ModelArch],
|
|
@@ -896,7 +898,7 @@ with demo:
|
|
| 896 |
hidden_leaderboard_table_for_search,
|
| 897 |
shown_columns,
|
| 898 |
search_bar,
|
| 899 |
-
filter_columns_type,
|
| 900 |
filter_domain_specific,
|
| 901 |
filter_columns_size
|
| 902 |
# filter_columns_architecture
|
|
@@ -905,7 +907,7 @@ with demo:
|
|
| 905 |
)
|
| 906 |
for selector in [
|
| 907 |
shown_columns,
|
| 908 |
-
filter_columns_type,
|
| 909 |
filter_domain_specific,
|
| 910 |
filter_columns_size,
|
| 911 |
# deleted_models_visibility,
|
|
@@ -916,7 +918,7 @@ with demo:
|
|
| 916 |
hidden_leaderboard_table_for_search,
|
| 917 |
shown_columns,
|
| 918 |
search_bar,
|
| 919 |
-
filter_columns_type,
|
| 920 |
filter_domain_specific,
|
| 921 |
filter_columns_size
|
| 922 |
],
|
|
@@ -950,13 +952,13 @@ with demo:
|
|
| 950 |
# )
|
| 951 |
with gr.Column(min_width=320):
|
| 952 |
# with gr.Box(elem_id="box-filter"):
|
| 953 |
-
filter_columns_type = gr.CheckboxGroup(
|
| 954 |
-
|
| 955 |
-
|
| 956 |
-
|
| 957 |
-
|
| 958 |
-
|
| 959 |
-
)
|
| 960 |
# filter_columns_architecture = gr.CheckboxGroup(
|
| 961 |
# label="Architecture Types",
|
| 962 |
# choices=[i.value.name for i in ModelArch],
|
|
@@ -1005,7 +1007,7 @@ with demo:
|
|
| 1005 |
hidden_leaderboard_table_for_search,
|
| 1006 |
shown_columns,
|
| 1007 |
search_bar,
|
| 1008 |
-
filter_columns_type,
|
| 1009 |
filter_domain_specific,
|
| 1010 |
filter_columns_size
|
| 1011 |
# filter_columns_architecture
|
|
@@ -1014,7 +1016,7 @@ with demo:
|
|
| 1014 |
)
|
| 1015 |
for selector in [
|
| 1016 |
shown_columns,
|
| 1017 |
-
filter_columns_type,
|
| 1018 |
filter_domain_specific,
|
| 1019 |
filter_columns_size,
|
| 1020 |
# deleted_models_visibility,
|
|
@@ -1025,7 +1027,7 @@ with demo:
|
|
| 1025 |
hidden_leaderboard_table_for_search,
|
| 1026 |
shown_columns,
|
| 1027 |
search_bar,
|
| 1028 |
-
filter_columns_type,
|
| 1029 |
filter_domain_specific,
|
| 1030 |
filter_columns_size
|
| 1031 |
],
|
|
@@ -1060,13 +1062,13 @@ with demo:
|
|
| 1060 |
# )
|
| 1061 |
with gr.Column(min_width=320):
|
| 1062 |
# with gr.Box(elem_id="box-filter"):
|
| 1063 |
-
filter_columns_type = gr.CheckboxGroup(
|
| 1064 |
-
|
| 1065 |
-
|
| 1066 |
-
|
| 1067 |
-
|
| 1068 |
-
|
| 1069 |
-
)
|
| 1070 |
# filter_columns_architecture = gr.CheckboxGroup(
|
| 1071 |
# label="Architecture Types",
|
| 1072 |
# choices=[i.value.name for i in ModelArch],
|
|
@@ -1115,7 +1117,7 @@ with demo:
|
|
| 1115 |
hidden_leaderboard_table_for_search,
|
| 1116 |
shown_columns,
|
| 1117 |
search_bar,
|
| 1118 |
-
filter_columns_type,
|
| 1119 |
filter_domain_specific,
|
| 1120 |
filter_columns_size
|
| 1121 |
# filter_columns_architecture
|
|
@@ -1124,7 +1126,7 @@ with demo:
|
|
| 1124 |
)
|
| 1125 |
for selector in [
|
| 1126 |
shown_columns,
|
| 1127 |
-
filter_columns_type,
|
| 1128 |
filter_domain_specific,
|
| 1129 |
filter_columns_size,
|
| 1130 |
# deleted_models_visibility,
|
|
@@ -1135,7 +1137,7 @@ with demo:
|
|
| 1135 |
hidden_leaderboard_table_for_search,
|
| 1136 |
shown_columns,
|
| 1137 |
search_bar,
|
| 1138 |
-
filter_columns_type,
|
| 1139 |
filter_domain_specific,
|
| 1140 |
filter_columns_size
|
| 1141 |
],
|
|
@@ -1173,13 +1175,13 @@ with demo:
|
|
| 1173 |
interactive=True,
|
| 1174 |
)
|
| 1175 |
with gr.Column(min_width=320):
|
| 1176 |
-
filter_columns_type = gr.CheckboxGroup(
|
| 1177 |
-
|
| 1178 |
-
|
| 1179 |
-
|
| 1180 |
-
|
| 1181 |
-
|
| 1182 |
-
)
|
| 1183 |
filter_domain_specific = gr.CheckboxGroup(
|
| 1184 |
label="Domain Specificity",
|
| 1185 |
choices=["🏥 Clinical models", "Generic models"],
|
|
@@ -1219,7 +1221,7 @@ with demo:
|
|
| 1219 |
hidden_leaderboard_table_for_search,
|
| 1220 |
shown_columns,
|
| 1221 |
search_bar,
|
| 1222 |
-
filter_columns_type,
|
| 1223 |
filter_domain_specific,
|
| 1224 |
filter_columns_size
|
| 1225 |
],
|
|
@@ -1228,7 +1230,7 @@ with demo:
|
|
| 1228 |
|
| 1229 |
for selector in [
|
| 1230 |
shown_columns,
|
| 1231 |
-
filter_columns_type,
|
| 1232 |
filter_domain_specific,
|
| 1233 |
filter_columns_size,
|
| 1234 |
]:
|
|
@@ -1238,7 +1240,7 @@ with demo:
|
|
| 1238 |
hidden_leaderboard_table_for_search,
|
| 1239 |
shown_columns,
|
| 1240 |
search_bar,
|
| 1241 |
-
filter_columns_type,
|
| 1242 |
filter_domain_specific,
|
| 1243 |
filter_columns_size
|
| 1244 |
],
|
|
@@ -1274,13 +1276,13 @@ with demo:
|
|
| 1274 |
)
|
| 1275 |
with gr.Column(min_width=320):
|
| 1276 |
# with gr.Box(elem_id="box-filter"):
|
| 1277 |
-
filter_columns_type = gr.CheckboxGroup(
|
| 1278 |
-
|
| 1279 |
-
|
| 1280 |
-
|
| 1281 |
-
|
| 1282 |
-
|
| 1283 |
-
)
|
| 1284 |
filter_domain_specific = gr.CheckboxGroup(
|
| 1285 |
label="Domain Specificity",
|
| 1286 |
choices=["🏥 Clinical models", "Generic models"],
|
|
@@ -1318,7 +1320,7 @@ with demo:
|
|
| 1318 |
hidden_leaderboard_table_for_search,
|
| 1319 |
shown_columns,
|
| 1320 |
search_bar,
|
| 1321 |
-
filter_columns_type,
|
| 1322 |
filter_domain_specific,
|
| 1323 |
filter_columns_size
|
| 1324 |
# filter_columns_architecture
|
|
@@ -1327,7 +1329,7 @@ with demo:
|
|
| 1327 |
)
|
| 1328 |
for selector in [
|
| 1329 |
shown_columns,
|
| 1330 |
-
filter_columns_type,
|
| 1331 |
filter_domain_specific,
|
| 1332 |
# filter_columns_architecture,
|
| 1333 |
filter_columns_size,
|
|
@@ -1339,7 +1341,7 @@ with demo:
|
|
| 1339 |
hidden_leaderboard_table_for_search,
|
| 1340 |
shown_columns,
|
| 1341 |
search_bar,
|
| 1342 |
-
filter_columns_type,
|
| 1343 |
filter_domain_specific,
|
| 1344 |
filter_columns_size
|
| 1345 |
# filter_columns_architecture,
|
|
|
|
| 239 |
hidden_df: pd.DataFrame,
|
| 240 |
columns: list,
|
| 241 |
query: str = "",
|
| 242 |
+
# type_query: list = None,
|
| 243 |
domain_specific_query: list = None,
|
| 244 |
size_query: list = None,
|
| 245 |
precision_query: str = None,
|
| 246 |
show_deleted: bool = False,
|
| 247 |
):
|
| 248 |
# breakpoint()
|
| 249 |
+
type_query = None
|
| 250 |
filtered_df = filter_models(hidden_df, type_query, domain_specific_query, size_query, precision_query, show_deleted)
|
| 251 |
# breakpoint()
|
| 252 |
filtered_df = filter_queries(query, filtered_df)
|
|
|
|
| 339 |
print("hello")
|
| 340 |
gr.HTML(LOGO)
|
| 341 |
gr.Markdown(INTRODUCTION_TEXT, elem_classes="markdown-text")
|
| 342 |
+
filter_columns_type = None
|
| 343 |
with gr.Tabs(elem_classes="tab-buttons") as tabs:
|
| 344 |
with gr.TabItem("🏅 Open Ended Evaluation", elem_id="llm-benchmark-tab-table", id=1):
|
| 345 |
with gr.Tabs(elem_classes="tab-buttons6") as language_tabs:
|
|
|
|
| 384 |
interactive=True,
|
| 385 |
)
|
| 386 |
with gr.Column(min_width=320):
|
| 387 |
+
# filter_columns_type = gr.CheckboxGroup(
|
| 388 |
+
# label="Model Types",
|
| 389 |
+
# choices=[t.to_str() for t in ModelType],
|
| 390 |
+
# value=[t.to_str() for t in ModelType],
|
| 391 |
+
# interactive=True,
|
| 392 |
+
# elem_id=f"filter-columns-type-{subset}",
|
| 393 |
+
# )
|
| 394 |
+
|
| 395 |
filter_domain_specific = gr.CheckboxGroup(
|
| 396 |
label="Domain Specificity",
|
| 397 |
choices=["🏥 Clinical models", "Generic models"],
|
|
|
|
| 431 |
hidden_leaderboard_table_for_search,
|
| 432 |
shown_columns,
|
| 433 |
search_bar,
|
| 434 |
+
# filter_columns_type,
|
| 435 |
filter_domain_specific,
|
| 436 |
filter_columns_size
|
| 437 |
],
|
|
|
|
| 440 |
|
| 441 |
for selector in [
|
| 442 |
shown_columns,
|
| 443 |
+
# filter_columns_type,
|
| 444 |
filter_domain_specific,
|
| 445 |
filter_columns_size,
|
| 446 |
]:
|
|
|
|
| 450 |
hidden_leaderboard_table_for_search,
|
| 451 |
shown_columns,
|
| 452 |
search_bar,
|
| 453 |
+
# filter_columns_type,
|
| 454 |
filter_domain_specific,
|
| 455 |
filter_columns_size
|
| 456 |
],
|
|
|
|
| 492 |
# )
|
| 493 |
with gr.Column(min_width=320):
|
| 494 |
# with gr.Box(elem_id="box-filter"):
|
| 495 |
+
# filter_columns_type = gr.CheckboxGroup(
|
| 496 |
+
# label="Model Types",
|
| 497 |
+
# choices=[t.to_str() for t in ModelType],
|
| 498 |
+
# value=[t.to_str() for t in ModelType],
|
| 499 |
+
# interactive=True,
|
| 500 |
+
# elem_id="filter-columns-type",
|
| 501 |
+
# )
|
| 502 |
# filter_columns_architecture = gr.CheckboxGroup(
|
| 503 |
# label="Architecture Types",
|
| 504 |
# choices=[i.value.name for i in ModelArch],
|
|
|
|
| 547 |
hidden_leaderboard_table_for_search,
|
| 548 |
shown_columns,
|
| 549 |
search_bar,
|
| 550 |
+
# filter_columns_type,
|
| 551 |
filter_domain_specific,
|
| 552 |
filter_columns_size
|
| 553 |
# filter_columns_architecture
|
|
|
|
| 556 |
)
|
| 557 |
for selector in [
|
| 558 |
shown_columns,
|
| 559 |
+
# filter_columns_type,
|
| 560 |
filter_domain_specific,
|
| 561 |
filter_columns_size,
|
| 562 |
# deleted_models_visibility,
|
|
|
|
| 567 |
hidden_leaderboard_table_for_search,
|
| 568 |
shown_columns,
|
| 569 |
search_bar,
|
| 570 |
+
# filter_columns_type,
|
| 571 |
filter_domain_specific,
|
| 572 |
filter_columns_size
|
| 573 |
],
|
|
|
|
| 612 |
# )
|
| 613 |
with gr.Column(min_width=320):
|
| 614 |
# with gr.Box(elem_id="box-filter"):
|
| 615 |
+
# filter_columns_type = gr.CheckboxGroup(
|
| 616 |
+
# label="Model Types",
|
| 617 |
+
# choices=[t.to_str() for t in ModelType],
|
| 618 |
+
# value=[t.to_str() for t in ModelType],
|
| 619 |
+
# interactive=True,
|
| 620 |
+
# elem_id="filter-columns-type",
|
| 621 |
+
# )
|
| 622 |
# filter_columns_architecture = gr.CheckboxGroup(
|
| 623 |
# label="Architecture Types",
|
| 624 |
# choices=[i.value.name for i in ModelArch],
|
|
|
|
| 667 |
hidden_leaderboard_table_for_search,
|
| 668 |
shown_columns,
|
| 669 |
search_bar,
|
| 670 |
+
# filter_columns_type,
|
| 671 |
filter_domain_specific,
|
| 672 |
filter_columns_size
|
| 673 |
# filter_columns_architecture
|
|
|
|
| 676 |
)
|
| 677 |
for selector in [
|
| 678 |
shown_columns,
|
| 679 |
+
# filter_columns_type,
|
| 680 |
filter_domain_specific,
|
| 681 |
filter_columns_size,
|
| 682 |
# deleted_models_visibility,
|
|
|
|
| 687 |
hidden_leaderboard_table_for_search,
|
| 688 |
shown_columns,
|
| 689 |
search_bar,
|
| 690 |
+
# filter_columns_type,
|
| 691 |
filter_domain_specific,
|
| 692 |
filter_columns_size
|
| 693 |
],
|
|
|
|
| 721 |
# )
|
| 722 |
with gr.Column(min_width=320):
|
| 723 |
# with gr.Box(elem_id="box-filter"):
|
| 724 |
+
# filter_columns_type = gr.CheckboxGroup(
|
| 725 |
+
# label="Model Types",
|
| 726 |
+
# choices=[t.to_str() for t in ModelType],
|
| 727 |
+
# value=[t.to_str() for t in ModelType],
|
| 728 |
+
# interactive=True,
|
| 729 |
+
# elem_id="filter-columns-type",
|
| 730 |
+
# )
|
| 731 |
# filter_columns_architecture = gr.CheckboxGroup(
|
| 732 |
# label="Architecture Types",
|
| 733 |
# choices=[i.value.name for i in ModelArch],
|
|
|
|
| 776 |
hidden_leaderboard_table_for_search,
|
| 777 |
shown_columns,
|
| 778 |
search_bar,
|
| 779 |
+
# filter_columns_type,
|
| 780 |
filter_domain_specific,
|
| 781 |
filter_columns_size
|
| 782 |
# filter_columns_architecture
|
|
|
|
| 785 |
)
|
| 786 |
for selector in [
|
| 787 |
shown_columns,
|
| 788 |
+
# filter_columns_type,
|
| 789 |
filter_domain_specific,
|
| 790 |
filter_columns_size,
|
| 791 |
# deleted_models_visibility,
|
|
|
|
| 796 |
hidden_leaderboard_table_for_search,
|
| 797 |
shown_columns,
|
| 798 |
search_bar,
|
| 799 |
+
# filter_columns_type,
|
| 800 |
filter_domain_specific,
|
| 801 |
filter_columns_size
|
| 802 |
],
|
|
|
|
| 843 |
# )
|
| 844 |
with gr.Column(min_width=320):
|
| 845 |
# with gr.Box(elem_id="box-filter"):
|
| 846 |
+
# filter_columns_type = gr.CheckboxGroup(
|
| 847 |
+
# label="Model Types",
|
| 848 |
+
# choices=[t.to_str() for t in ModelType],
|
| 849 |
+
# value=[t.to_str() for t in ModelType],
|
| 850 |
+
# interactive=True,
|
| 851 |
+
# elem_id="filter-columns-type",
|
| 852 |
+
# )
|
| 853 |
# filter_columns_architecture = gr.CheckboxGroup(
|
| 854 |
# label="Architecture Types",
|
| 855 |
# choices=[i.value.name for i in ModelArch],
|
|
|
|
| 898 |
hidden_leaderboard_table_for_search,
|
| 899 |
shown_columns,
|
| 900 |
search_bar,
|
| 901 |
+
# filter_columns_type,
|
| 902 |
filter_domain_specific,
|
| 903 |
filter_columns_size
|
| 904 |
# filter_columns_architecture
|
|
|
|
| 907 |
)
|
| 908 |
for selector in [
|
| 909 |
shown_columns,
|
| 910 |
+
# filter_columns_type,
|
| 911 |
filter_domain_specific,
|
| 912 |
filter_columns_size,
|
| 913 |
# deleted_models_visibility,
|
|
|
|
| 918 |
hidden_leaderboard_table_for_search,
|
| 919 |
shown_columns,
|
| 920 |
search_bar,
|
| 921 |
+
# filter_columns_type,
|
| 922 |
filter_domain_specific,
|
| 923 |
filter_columns_size
|
| 924 |
],
|
|
|
|
| 952 |
# )
|
| 953 |
with gr.Column(min_width=320):
|
| 954 |
# with gr.Box(elem_id="box-filter"):
|
| 955 |
+
# filter_columns_type = gr.CheckboxGroup(
|
| 956 |
+
# label="Model Types",
|
| 957 |
+
# choices=[t.to_str() for t in ModelType],
|
| 958 |
+
# value=[t.to_str() for t in ModelType],
|
| 959 |
+
# interactive=True,
|
| 960 |
+
# elem_id="filter-columns-type",
|
| 961 |
+
# )
|
| 962 |
# filter_columns_architecture = gr.CheckboxGroup(
|
| 963 |
# label="Architecture Types",
|
| 964 |
# choices=[i.value.name for i in ModelArch],
|
|
|
|
| 1007 |
hidden_leaderboard_table_for_search,
|
| 1008 |
shown_columns,
|
| 1009 |
search_bar,
|
| 1010 |
+
# filter_columns_type,
|
| 1011 |
filter_domain_specific,
|
| 1012 |
filter_columns_size
|
| 1013 |
# filter_columns_architecture
|
|
|
|
| 1016 |
)
|
| 1017 |
for selector in [
|
| 1018 |
shown_columns,
|
| 1019 |
+
# filter_columns_type,
|
| 1020 |
filter_domain_specific,
|
| 1021 |
filter_columns_size,
|
| 1022 |
# deleted_models_visibility,
|
|
|
|
| 1027 |
hidden_leaderboard_table_for_search,
|
| 1028 |
shown_columns,
|
| 1029 |
search_bar,
|
| 1030 |
+
# filter_columns_type,
|
| 1031 |
filter_domain_specific,
|
| 1032 |
filter_columns_size
|
| 1033 |
],
|
|
|
|
| 1062 |
# )
|
| 1063 |
with gr.Column(min_width=320):
|
| 1064 |
# with gr.Box(elem_id="box-filter"):
|
| 1065 |
+
# filter_columns_type = gr.CheckboxGroup(
|
| 1066 |
+
# label="Model Types",
|
| 1067 |
+
# choices=[t.to_str() for t in ModelType],
|
| 1068 |
+
# value=[t.to_str() for t in ModelType],
|
| 1069 |
+
# interactive=True,
|
| 1070 |
+
# elem_id="filter-columns-type",
|
| 1071 |
+
# )
|
| 1072 |
# filter_columns_architecture = gr.CheckboxGroup(
|
| 1073 |
# label="Architecture Types",
|
| 1074 |
# choices=[i.value.name for i in ModelArch],
|
|
|
|
| 1117 |
hidden_leaderboard_table_for_search,
|
| 1118 |
shown_columns,
|
| 1119 |
search_bar,
|
| 1120 |
+
# filter_columns_type,
|
| 1121 |
filter_domain_specific,
|
| 1122 |
filter_columns_size
|
| 1123 |
# filter_columns_architecture
|
|
|
|
| 1126 |
)
|
| 1127 |
for selector in [
|
| 1128 |
shown_columns,
|
| 1129 |
+
# filter_columns_type,
|
| 1130 |
filter_domain_specific,
|
| 1131 |
filter_columns_size,
|
| 1132 |
# deleted_models_visibility,
|
|
|
|
| 1137 |
hidden_leaderboard_table_for_search,
|
| 1138 |
shown_columns,
|
| 1139 |
search_bar,
|
| 1140 |
+
# filter_columns_type,
|
| 1141 |
filter_domain_specific,
|
| 1142 |
filter_columns_size
|
| 1143 |
],
|
|
|
|
| 1175 |
interactive=True,
|
| 1176 |
)
|
| 1177 |
with gr.Column(min_width=320):
|
| 1178 |
+
# filter_columns_type = gr.CheckboxGroup(
|
| 1179 |
+
# label="Model Types",
|
| 1180 |
+
# choices=[t.to_str() for t in ModelType],
|
| 1181 |
+
# value=[t.to_str() for t in ModelType],
|
| 1182 |
+
# interactive=True,
|
| 1183 |
+
# elem_id="filter-columns-type-closed-english",
|
| 1184 |
+
# )
|
| 1185 |
filter_domain_specific = gr.CheckboxGroup(
|
| 1186 |
label="Domain Specificity",
|
| 1187 |
choices=["🏥 Clinical models", "Generic models"],
|
|
|
|
| 1221 |
hidden_leaderboard_table_for_search,
|
| 1222 |
shown_columns,
|
| 1223 |
search_bar,
|
| 1224 |
+
# filter_columns_type,
|
| 1225 |
filter_domain_specific,
|
| 1226 |
filter_columns_size
|
| 1227 |
],
|
|
|
|
| 1230 |
|
| 1231 |
for selector in [
|
| 1232 |
shown_columns,
|
| 1233 |
+
# filter_columns_type,
|
| 1234 |
filter_domain_specific,
|
| 1235 |
filter_columns_size,
|
| 1236 |
]:
|
|
|
|
| 1240 |
hidden_leaderboard_table_for_search,
|
| 1241 |
shown_columns,
|
| 1242 |
search_bar,
|
| 1243 |
+
# filter_columns_type,
|
| 1244 |
filter_domain_specific,
|
| 1245 |
filter_columns_size
|
| 1246 |
],
|
|
|
|
| 1276 |
)
|
| 1277 |
with gr.Column(min_width=320):
|
| 1278 |
# with gr.Box(elem_id="box-filter"):
|
| 1279 |
+
# filter_columns_type = gr.CheckboxGroup(
|
| 1280 |
+
# label="Model Types",
|
| 1281 |
+
# choices=[t.to_str() for t in ModelType],
|
| 1282 |
+
# value=[t.to_str() for t in ModelType],
|
| 1283 |
+
# interactive=True,
|
| 1284 |
+
# elem_id="filter-columns-type",
|
| 1285 |
+
# )
|
| 1286 |
filter_domain_specific = gr.CheckboxGroup(
|
| 1287 |
label="Domain Specificity",
|
| 1288 |
choices=["🏥 Clinical models", "Generic models"],
|
|
|
|
| 1320 |
hidden_leaderboard_table_for_search,
|
| 1321 |
shown_columns,
|
| 1322 |
search_bar,
|
| 1323 |
+
# filter_columns_type,
|
| 1324 |
filter_domain_specific,
|
| 1325 |
filter_columns_size
|
| 1326 |
# filter_columns_architecture
|
|
|
|
| 1329 |
)
|
| 1330 |
for selector in [
|
| 1331 |
shown_columns,
|
| 1332 |
+
# filter_columns_type,
|
| 1333 |
filter_domain_specific,
|
| 1334 |
# filter_columns_architecture,
|
| 1335 |
filter_columns_size,
|
|
|
|
| 1341 |
hidden_leaderboard_table_for_search,
|
| 1342 |
shown_columns,
|
| 1343 |
search_bar,
|
| 1344 |
+
# filter_columns_type,
|
| 1345 |
filter_domain_specific,
|
| 1346 |
filter_columns_size
|
| 1347 |
# filter_columns_architecture,
|
src/display/utils.py
CHANGED
|
@@ -46,7 +46,7 @@ class ColumnContent:
|
|
| 46 |
## Leaderboard columns
|
| 47 |
# Init
|
| 48 |
auto_eval_column_dict = []
|
| 49 |
-
auto_eval_column_dict.append(["model_type_symbol", ColumnContent, ColumnContent("T", "str", True, never_hidden=True)])
|
| 50 |
auto_eval_column_dict.append(["model", ColumnContent, ColumnContent("Model", "markdown", True, never_hidden=True)])
|
| 51 |
auto_eval_column_dict.append(["revision", ColumnContent, ColumnContent("Model sha", "str", False, True)])
|
| 52 |
auto_eval_column_dict.append(["average", ColumnContent, ColumnContent("Average", "number", True, False, dataset_task_col=True, closed_ended_multilingual_col=True, invariant=False)])
|
|
|
|
| 46 |
## Leaderboard columns
|
| 47 |
# Init
|
| 48 |
auto_eval_column_dict = []
|
| 49 |
+
# auto_eval_column_dict.append(["model_type_symbol", ColumnContent, ColumnContent("T", "str", True, never_hidden=True)])
|
| 50 |
auto_eval_column_dict.append(["model", ColumnContent, ColumnContent("Model", "markdown", True, never_hidden=True)])
|
| 51 |
auto_eval_column_dict.append(["revision", ColumnContent, ColumnContent("Model sha", "str", False, True)])
|
| 52 |
auto_eval_column_dict.append(["average", ColumnContent, ColumnContent("Average", "number", True, False, dataset_task_col=True, closed_ended_multilingual_col=True, invariant=False)])
|
src/leaderboard/read_evals.py
CHANGED
|
@@ -368,7 +368,7 @@ class EvalResult:
|
|
| 368 |
"eval_name": self.eval_name, # not a column, just a save name,
|
| 369 |
AutoEvalColumn.precision.name: self.precision.value.name,
|
| 370 |
AutoEvalColumn.model_type.name: self.model_type.value.name,
|
| 371 |
-
AutoEvalColumn.model_type_symbol.name: self.model_type.value.symbol + (" 🏥" if self.is_domain_specific else ""),
|
| 372 |
AutoEvalColumn.weight_type.name: self.weight_type.value.name,
|
| 373 |
# AutoEvalColumn.architecture.name: self.architecture.value.name,
|
| 374 |
# AutoEvalColumn.backbone.name: self.backbone,
|
|
|
|
| 368 |
"eval_name": self.eval_name, # not a column, just a save name,
|
| 369 |
AutoEvalColumn.precision.name: self.precision.value.name,
|
| 370 |
AutoEvalColumn.model_type.name: self.model_type.value.name,
|
| 371 |
+
# AutoEvalColumn.model_type_symbol.name: self.model_type.value.symbol + (" 🏥" if self.is_domain_specific else ""),
|
| 372 |
AutoEvalColumn.weight_type.name: self.weight_type.value.name,
|
| 373 |
# AutoEvalColumn.architecture.name: self.architecture.value.name,
|
| 374 |
# AutoEvalColumn.backbone.name: self.backbone,
|
src/populate.py
CHANGED
|
@@ -16,7 +16,7 @@ def get_leaderboard_df(results_path: str, requests_path: str, cols: list, benchm
|
|
| 16 |
# raise Exception("stop")
|
| 17 |
# if subset.startswith("healthbench"):
|
| 18 |
# breakpoint()
|
| 19 |
-
all_data_json = [v.to_dict(subset=subset) for v in raw_data]
|
| 20 |
df = pd.DataFrame.from_records(all_data_json)
|
| 21 |
# changes to be made here
|
| 22 |
if subset == "datasets":
|
|
|
|
| 16 |
# raise Exception("stop")
|
| 17 |
# if subset.startswith("healthbench"):
|
| 18 |
# breakpoint()
|
| 19 |
+
all_data_json = [v.to_dict(subset=subset) for v in raw_data if not v.full_model.startswith("models_llm")]
|
| 20 |
df = pd.DataFrame.from_records(all_data_json)
|
| 21 |
# changes to be made here
|
| 22 |
if subset == "datasets":
|