Spaces:
Running
on
CPU Upgrade
Running
on
CPU Upgrade
feat: add toggle checkbox for long-doc
Browse files- app.py +49 -45
- src/utils.py +9 -5
app.py
CHANGED
|
@@ -51,6 +51,7 @@ leaderboard_df_long_doc = original_df_long_doc.copy()
|
|
| 51 |
shown_columns_long_doc, types_long_doc = get_default_cols(
|
| 52 |
'long-doc', leaderboard_df_long_doc.columns, add_fix_cols=True)
|
| 53 |
leaderboard_df_long_doc = leaderboard_df_long_doc[~leaderboard_df_long_doc[COL_NAME_IS_ANONYMOUS]][shown_columns_long_doc]
|
|
|
|
| 54 |
|
| 55 |
|
| 56 |
def update_metric_qa(
|
|
@@ -59,9 +60,10 @@ def update_metric_qa(
|
|
| 59 |
langs: list,
|
| 60 |
reranking_model: list,
|
| 61 |
query: str,
|
| 62 |
-
show_anonymous: bool
|
|
|
|
| 63 |
):
|
| 64 |
-
return update_metric(raw_data, 'qa', metric, domains, langs, reranking_model, query, show_anonymous)
|
| 65 |
|
| 66 |
|
| 67 |
def update_metric_long_doc(
|
|
@@ -70,9 +72,10 @@ def update_metric_long_doc(
|
|
| 70 |
langs: list,
|
| 71 |
reranking_model: list,
|
| 72 |
query: str,
|
| 73 |
-
show_anonymous: bool
|
|
|
|
| 74 |
):
|
| 75 |
-
return update_metric(raw_data, "long-doc", metric, domains, langs, reranking_model, query, show_anonymous)
|
| 76 |
|
| 77 |
|
| 78 |
def update_table_without_ranking(
|
|
@@ -93,9 +96,10 @@ def update_table_without_ranking_long_doc(
|
|
| 93 |
langs,
|
| 94 |
reranking_query,
|
| 95 |
query,
|
| 96 |
-
show_anonymous
|
|
|
|
| 97 |
):
|
| 98 |
-
return update_table_long_doc(hidden_df, domains, langs, reranking_query, query, show_anonymous, reset_ranking=False)
|
| 99 |
|
| 100 |
|
| 101 |
demo = gr.Blocks(css=custom_css)
|
|
@@ -195,33 +199,21 @@ with demo:
|
|
| 195 |
leaderboard_table,
|
| 196 |
)
|
| 197 |
|
| 198 |
-
show_revision_and_timestamp
|
| 199 |
-
|
| 200 |
-
|
| 201 |
-
|
| 202 |
-
|
| 203 |
-
|
| 204 |
-
|
| 205 |
-
|
| 206 |
-
|
| 207 |
-
|
| 208 |
-
|
| 209 |
-
|
| 210 |
-
|
| 211 |
-
|
| 212 |
-
|
| 213 |
-
update_table_without_ranking,
|
| 214 |
-
[
|
| 215 |
-
hidden_leaderboard_table_for_search,
|
| 216 |
-
selected_domains,
|
| 217 |
-
selected_langs,
|
| 218 |
-
selected_rerankings,
|
| 219 |
-
search_bar,
|
| 220 |
-
show_anonymous,
|
| 221 |
-
show_revision_and_timestamp
|
| 222 |
-
],
|
| 223 |
-
leaderboard_table,
|
| 224 |
-
)
|
| 225 |
|
| 226 |
# Set column-wise listener
|
| 227 |
for selector in [
|
|
@@ -311,6 +303,12 @@ with demo:
|
|
| 311 |
value=False,
|
| 312 |
info="The anonymous submissions might have invalid model information."
|
| 313 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 314 |
|
| 315 |
leaderboard_table_long_doc = gr.components.Dataframe(
|
| 316 |
value=leaderboard_df_long_doc,
|
|
@@ -337,22 +335,26 @@ with demo:
|
|
| 337 |
selected_rerankings,
|
| 338 |
search_bar,
|
| 339 |
show_anonymous,
|
|
|
|
| 340 |
],
|
| 341 |
leaderboard_table_long_doc,
|
| 342 |
)
|
| 343 |
|
| 344 |
-
selected_rerankings
|
| 345 |
-
|
| 346 |
-
|
| 347 |
-
|
| 348 |
-
|
| 349 |
-
|
| 350 |
-
|
| 351 |
-
|
| 352 |
-
|
| 353 |
-
|
| 354 |
-
|
| 355 |
-
|
|
|
|
|
|
|
|
|
|
| 356 |
|
| 357 |
# Set column-wise listener
|
| 358 |
for selector in [
|
|
@@ -367,6 +369,7 @@ with demo:
|
|
| 367 |
selected_rerankings,
|
| 368 |
search_bar,
|
| 369 |
show_anonymous,
|
|
|
|
| 370 |
],
|
| 371 |
leaderboard_table_long_doc,
|
| 372 |
queue=True,
|
|
@@ -382,6 +385,7 @@ with demo:
|
|
| 382 |
selected_rerankings,
|
| 383 |
search_bar,
|
| 384 |
show_anonymous,
|
|
|
|
| 385 |
],
|
| 386 |
leaderboard_table_long_doc,
|
| 387 |
queue=True
|
|
|
|
| 51 |
shown_columns_long_doc, types_long_doc = get_default_cols(
|
| 52 |
'long-doc', leaderboard_df_long_doc.columns, add_fix_cols=True)
|
| 53 |
leaderboard_df_long_doc = leaderboard_df_long_doc[~leaderboard_df_long_doc[COL_NAME_IS_ANONYMOUS]][shown_columns_long_doc]
|
| 54 |
+
leaderboard_df_long_doc.drop([COL_NAME_REVISION, COL_NAME_TIMESTAMP], axis=1, inplace=True)
|
| 55 |
|
| 56 |
|
| 57 |
def update_metric_qa(
|
|
|
|
| 60 |
langs: list,
|
| 61 |
reranking_model: list,
|
| 62 |
query: str,
|
| 63 |
+
show_anonymous: bool,
|
| 64 |
+
show_revision_and_timestamp,
|
| 65 |
):
|
| 66 |
+
return update_metric(raw_data, 'qa', metric, domains, langs, reranking_model, query, show_anonymous, show_revision_and_timestamp)
|
| 67 |
|
| 68 |
|
| 69 |
def update_metric_long_doc(
|
|
|
|
| 72 |
langs: list,
|
| 73 |
reranking_model: list,
|
| 74 |
query: str,
|
| 75 |
+
show_anonymous: bool,
|
| 76 |
+
show_revision_and_timestamp,
|
| 77 |
):
|
| 78 |
+
return update_metric(raw_data, "long-doc", metric, domains, langs, reranking_model, query, show_anonymous, show_revision_and_timestamp)
|
| 79 |
|
| 80 |
|
| 81 |
def update_table_without_ranking(
|
|
|
|
| 96 |
langs,
|
| 97 |
reranking_query,
|
| 98 |
query,
|
| 99 |
+
show_anonymous,
|
| 100 |
+
show_revision_and_timestamp,
|
| 101 |
):
|
| 102 |
+
return update_table_long_doc(hidden_df, domains, langs, reranking_query, query, show_anonymous, reset_ranking=False, show_revision_and_timestamp=show_revision_and_timestamp)
|
| 103 |
|
| 104 |
|
| 105 |
demo = gr.Blocks(css=custom_css)
|
|
|
|
| 199 |
leaderboard_table,
|
| 200 |
)
|
| 201 |
|
| 202 |
+
for selector in [show_revision_and_timestamp, selected_rerankings]:
|
| 203 |
+
selector.change(
|
| 204 |
+
update_table_without_ranking,
|
| 205 |
+
[
|
| 206 |
+
hidden_leaderboard_table_for_search,
|
| 207 |
+
selected_domains,
|
| 208 |
+
selected_langs,
|
| 209 |
+
selected_rerankings,
|
| 210 |
+
search_bar,
|
| 211 |
+
show_anonymous,
|
| 212 |
+
show_revision_and_timestamp
|
| 213 |
+
],
|
| 214 |
+
leaderboard_table,
|
| 215 |
+
queue=True
|
| 216 |
+
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 217 |
|
| 218 |
# Set column-wise listener
|
| 219 |
for selector in [
|
|
|
|
| 303 |
value=False,
|
| 304 |
info="The anonymous submissions might have invalid model information."
|
| 305 |
)
|
| 306 |
+
with gr.Row():
|
| 307 |
+
show_revision_and_timestamp = gr.Checkbox(
|
| 308 |
+
label="Show submission details",
|
| 309 |
+
value=False,
|
| 310 |
+
info="Show the revision and timestamp information of submissions"
|
| 311 |
+
)
|
| 312 |
|
| 313 |
leaderboard_table_long_doc = gr.components.Dataframe(
|
| 314 |
value=leaderboard_df_long_doc,
|
|
|
|
| 335 |
selected_rerankings,
|
| 336 |
search_bar,
|
| 337 |
show_anonymous,
|
| 338 |
+
show_revision_and_timestamp
|
| 339 |
],
|
| 340 |
leaderboard_table_long_doc,
|
| 341 |
)
|
| 342 |
|
| 343 |
+
for selector in [show_revision_and_timestamp, selected_rerankings]:
|
| 344 |
+
selector.change(
|
| 345 |
+
update_table_without_ranking_long_doc,
|
| 346 |
+
[
|
| 347 |
+
hidden_leaderboard_table_for_search,
|
| 348 |
+
selected_domains,
|
| 349 |
+
selected_langs,
|
| 350 |
+
selected_rerankings,
|
| 351 |
+
search_bar,
|
| 352 |
+
show_anonymous,
|
| 353 |
+
show_revision_and_timestamp
|
| 354 |
+
],
|
| 355 |
+
leaderboard_table_long_doc,
|
| 356 |
+
queue=True,
|
| 357 |
+
)
|
| 358 |
|
| 359 |
# Set column-wise listener
|
| 360 |
for selector in [
|
|
|
|
| 369 |
selected_rerankings,
|
| 370 |
search_bar,
|
| 371 |
show_anonymous,
|
| 372 |
+
show_revision_and_timestamp
|
| 373 |
],
|
| 374 |
leaderboard_table_long_doc,
|
| 375 |
queue=True,
|
|
|
|
| 385 |
selected_rerankings,
|
| 386 |
search_bar,
|
| 387 |
show_anonymous,
|
| 388 |
+
show_revision_and_timestamp
|
| 389 |
],
|
| 390 |
leaderboard_table_long_doc,
|
| 391 |
queue=True
|
src/utils.py
CHANGED
|
@@ -165,10 +165,11 @@ def update_table_long_doc(
|
|
| 165 |
reranking_query: list,
|
| 166 |
query: str,
|
| 167 |
show_anonymous: bool,
|
| 168 |
-
reset_ranking: bool = True
|
|
|
|
| 169 |
):
|
| 170 |
return _update_table(
|
| 171 |
-
"long-doc", hidden_df, domains, langs, reranking_query, query, show_anonymous, reset_ranking)
|
| 172 |
|
| 173 |
|
| 174 |
def update_metric(
|
|
@@ -179,7 +180,8 @@ def update_metric(
|
|
| 179 |
langs: list,
|
| 180 |
reranking_model: list,
|
| 181 |
query: str,
|
| 182 |
-
show_anonymous: bool = False
|
|
|
|
| 183 |
) -> pd.DataFrame:
|
| 184 |
if task == 'qa':
|
| 185 |
leaderboard_df = get_leaderboard_df(raw_data, task=task, metric=metric)
|
|
@@ -189,7 +191,8 @@ def update_metric(
|
|
| 189 |
langs,
|
| 190 |
reranking_model,
|
| 191 |
query,
|
| 192 |
-
show_anonymous
|
|
|
|
| 193 |
)
|
| 194 |
elif task == "long-doc":
|
| 195 |
leaderboard_df = get_leaderboard_df(raw_data, task=task, metric=metric)
|
|
@@ -199,7 +202,8 @@ def update_metric(
|
|
| 199 |
langs,
|
| 200 |
reranking_model,
|
| 201 |
query,
|
| 202 |
-
show_anonymous
|
|
|
|
| 203 |
)
|
| 204 |
|
| 205 |
|
|
|
|
| 165 |
reranking_query: list,
|
| 166 |
query: str,
|
| 167 |
show_anonymous: bool,
|
| 168 |
+
reset_ranking: bool = True,
|
| 169 |
+
show_revision_and_timestamp: bool = False
|
| 170 |
):
|
| 171 |
return _update_table(
|
| 172 |
+
"long-doc", hidden_df, domains, langs, reranking_query, query, show_anonymous, reset_ranking, show_revision_and_timestamp)
|
| 173 |
|
| 174 |
|
| 175 |
def update_metric(
|
|
|
|
| 180 |
langs: list,
|
| 181 |
reranking_model: list,
|
| 182 |
query: str,
|
| 183 |
+
show_anonymous: bool = False,
|
| 184 |
+
show_revision_and_timestamp: bool = False,
|
| 185 |
) -> pd.DataFrame:
|
| 186 |
if task == 'qa':
|
| 187 |
leaderboard_df = get_leaderboard_df(raw_data, task=task, metric=metric)
|
|
|
|
| 191 |
langs,
|
| 192 |
reranking_model,
|
| 193 |
query,
|
| 194 |
+
show_anonymous,
|
| 195 |
+
show_revision_and_timestamp
|
| 196 |
)
|
| 197 |
elif task == "long-doc":
|
| 198 |
leaderboard_df = get_leaderboard_df(raw_data, task=task, metric=metric)
|
|
|
|
| 202 |
langs,
|
| 203 |
reranking_model,
|
| 204 |
query,
|
| 205 |
+
show_anonymous,
|
| 206 |
+
show_revision_and_timestamp
|
| 207 |
)
|
| 208 |
|
| 209 |
|