Spaces:
Running
on
CPU Upgrade
Running
on
CPU Upgrade
fix: fix the qa version updating bug
Browse files
app.py
CHANGED
|
@@ -107,17 +107,87 @@ def update_doc_metric(
|
|
| 107 |
show_revision_and_timestamp,
|
| 108 |
)
|
| 109 |
|
| 110 |
-
|
| 111 |
-
def update_qa_version(version):
|
| 112 |
global datastore
|
| 113 |
global ds_dict
|
| 114 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 115 |
domain_elem = get_domain_dropdown(QABenchmarks[datastore.slug])
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 116 |
lang_elem = get_language_dropdown(QABenchmarks[datastore.slug])
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 117 |
model_elem = get_reranking_dropdown(datastore.reranking_models)
|
|
|
|
|
|
|
|
|
|
|
|
|
| 118 |
df_elem = get_leaderboard_table(datastore.qa_fmt_df, datastore.qa_types)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 119 |
hidden_df_elem = get_leaderboard_table(datastore.qa_raw_df, datastore.qa_types, visible=False)
|
| 120 |
-
return
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 121 |
|
| 122 |
|
| 123 |
def update_doc_version(version):
|
|
@@ -151,9 +221,19 @@ with demo:
|
|
| 151 |
# select domain
|
| 152 |
with gr.Row():
|
| 153 |
domains = get_domain_dropdown(QABenchmarks[datastore.slug])
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 154 |
# select language
|
| 155 |
with gr.Row():
|
| 156 |
langs = get_language_dropdown(QABenchmarks[datastore.slug])
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 157 |
with gr.Column():
|
| 158 |
# select the metric
|
| 159 |
metric = get_metric_dropdown(METRIC_LIST, DEFAULT_METRIC_QA)
|
|
@@ -170,17 +250,26 @@ with demo:
|
|
| 170 |
# select reranking models
|
| 171 |
with gr.Column():
|
| 172 |
models = get_reranking_dropdown(datastore.reranking_models)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 173 |
# shown_table
|
| 174 |
qa_df_elem_ret_rerank = get_leaderboard_table(datastore.qa_fmt_df, datastore.qa_types)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 175 |
# Dummy leaderboard for handling the case when the user uses backspace key
|
| 176 |
qa_df_elem_ret_rerank_hidden = get_leaderboard_table(
|
| 177 |
datastore.qa_raw_df, datastore.qa_types, visible=False
|
| 178 |
)
|
| 179 |
-
|
| 180 |
version.change(
|
| 181 |
-
|
| 182 |
version,
|
| 183 |
-
|
| 184 |
)
|
| 185 |
|
| 186 |
set_listeners(
|
|
@@ -210,10 +299,21 @@ with demo:
|
|
| 210 |
search_bar_ret = get_search_bar()
|
| 211 |
with gr.Column(scale=1):
|
| 212 |
models_ret = get_noreranking_dropdown()
|
| 213 |
-
|
| 214 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 215 |
_qa_df_ret = reset_rank(_qa_df_ret)
|
| 216 |
qa_df_elem_ret = get_leaderboard_table(_qa_df_ret, datastore.qa_types)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 217 |
|
| 218 |
# Dummy leaderboard for handling the case when the user uses backspace key
|
| 219 |
_qa_df_ret_hidden = datastore.qa_raw_df[
|
|
@@ -225,15 +325,9 @@ with demo:
|
|
| 225 |
)
|
| 226 |
|
| 227 |
version.change(
|
| 228 |
-
|
| 229 |
version,
|
| 230 |
-
|
| 231 |
-
domains,
|
| 232 |
-
langs,
|
| 233 |
-
models_ret,
|
| 234 |
-
qa_df_elem_ret,
|
| 235 |
-
qa_df_elem_ret_hidden,
|
| 236 |
-
],
|
| 237 |
)
|
| 238 |
|
| 239 |
set_listeners(
|
|
@@ -271,9 +365,19 @@ with demo:
|
|
| 271 |
with gr.Row():
|
| 272 |
with gr.Column(scale=1):
|
| 273 |
qa_models_rerank = get_reranking_dropdown(qa_rerank_models)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 274 |
with gr.Column(scale=1):
|
| 275 |
qa_search_bar_rerank = gr.Textbox(show_label=False, visible=False)
|
| 276 |
qa_df_elem_rerank = get_leaderboard_table(_qa_df_rerank, datastore.qa_types)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 277 |
|
| 278 |
_qa_df_rerank_hidden = datastore.qa_raw_df[
|
| 279 |
datastore.qa_raw_df[COL_NAME_RETRIEVAL_MODEL] == BM25_LINK
|
|
@@ -282,11 +386,10 @@ with demo:
|
|
| 282 |
qa_df_elem_rerank_hidden = get_leaderboard_table(
|
| 283 |
_qa_df_rerank_hidden, datastore.qa_types, visible=False
|
| 284 |
)
|
| 285 |
-
|
| 286 |
version.change(
|
| 287 |
-
|
| 288 |
version,
|
| 289 |
-
|
| 290 |
)
|
| 291 |
|
| 292 |
set_listeners(
|
|
@@ -406,7 +509,13 @@ with demo:
|
|
| 406 |
version.change(
|
| 407 |
update_doc_version,
|
| 408 |
version,
|
| 409 |
-
[
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 410 |
)
|
| 411 |
|
| 412 |
set_listeners(
|
|
|
|
| 107 |
show_revision_and_timestamp,
|
| 108 |
)
|
| 109 |
|
| 110 |
+
def update_datastore(version):
|
|
|
|
| 111 |
global datastore
|
| 112 |
global ds_dict
|
| 113 |
+
print(f"current data version: {datastore.version}")
|
| 114 |
+
if datastore.version != version:
|
| 115 |
+
print(f"reload data version: {version}")
|
| 116 |
+
datastore = ds_dict[version]
|
| 117 |
+
return datastore
|
| 118 |
+
|
| 119 |
+
def update_qa_domains(version):
|
| 120 |
+
datastore = update_datastore(version)
|
| 121 |
domain_elem = get_domain_dropdown(QABenchmarks[datastore.slug])
|
| 122 |
+
return domain_elem
|
| 123 |
+
|
| 124 |
+
|
| 125 |
+
def update_qa_langs(version):
|
| 126 |
+
datastore = update_datastore(version)
|
| 127 |
lang_elem = get_language_dropdown(QABenchmarks[datastore.slug])
|
| 128 |
+
return lang_elem
|
| 129 |
+
|
| 130 |
+
|
| 131 |
+
def update_qa_models(version):
|
| 132 |
+
datastore = update_datastore(version)
|
| 133 |
model_elem = get_reranking_dropdown(datastore.reranking_models)
|
| 134 |
+
return model_elem
|
| 135 |
+
|
| 136 |
+
def update_qa_df_ret_rerank(version):
|
| 137 |
+
datastore = update_datastore(version)
|
| 138 |
df_elem = get_leaderboard_table(datastore.qa_fmt_df, datastore.qa_types)
|
| 139 |
+
return df_elem
|
| 140 |
+
|
| 141 |
+
def update_qa_df_ret(version):
|
| 142 |
+
datastore = update_datastore(version)
|
| 143 |
+
_qa_df_ret = datastore.qa_fmt_df[
|
| 144 |
+
datastore.qa_fmt_df[COL_NAME_RERANKING_MODEL] == "NoReranker"]
|
| 145 |
+
_qa_df_ret = reset_rank(_qa_df_ret)
|
| 146 |
+
df_elem_ret = get_leaderboard_table(_qa_df_ret, datastore.qa_types)
|
| 147 |
+
return df_elem_ret
|
| 148 |
+
|
| 149 |
+
def update_qa_df_rerank(version):
|
| 150 |
+
datastore = update_datastore(version)
|
| 151 |
+
_qa_df_rerank = datastore.qa_fmt_df[datastore.qa_fmt_df[COL_NAME_RETRIEVAL_MODEL] == BM25_LINK]
|
| 152 |
+
_qa_df_rerank = reset_rank(_qa_df_rerank)
|
| 153 |
+
df_elem_ret = get_leaderboard_table(_qa_df_rerank, datastore.qa_types)
|
| 154 |
+
return df_elem_ret
|
| 155 |
+
|
| 156 |
+
def update_qa_hidden_df_rerank(version):
|
| 157 |
+
datastore = update_datastore(version)
|
| 158 |
+
_qa_df_rerank_hidden = datastore.qa_raw_df[
|
| 159 |
+
datastore.qa_raw_df[COL_NAME_RETRIEVAL_MODEL] == BM25_LINK
|
| 160 |
+
]
|
| 161 |
+
_qa_df_rerank_hidden = reset_rank(_qa_df_rerank_hidden)
|
| 162 |
+
hidden_df_elem = get_leaderboard_table(
|
| 163 |
+
_qa_df_rerank_hidden, datastore.qa_types, visible=False
|
| 164 |
+
)
|
| 165 |
+
return hidden_df_elem
|
| 166 |
+
|
| 167 |
+
def update_qa_hidden_df_ret(version):
|
| 168 |
+
datastore = update_datastore(version)
|
| 169 |
+
_qa_df_ret_hidden = datastore.qa_raw_df[
|
| 170 |
+
datastore.qa_raw_df[COL_NAME_RERANKING_MODEL] == "NoReranker"
|
| 171 |
+
]
|
| 172 |
+
_qa_df_ret_hidden = reset_rank(_qa_df_ret_hidden)
|
| 173 |
+
hidden_df_elem = get_leaderboard_table(_qa_df_ret_hidden, datastore.qa_types, visible=False)
|
| 174 |
+
return hidden_df_elem
|
| 175 |
+
|
| 176 |
+
def update_qa_hidden_df_ret_rerank(version):
|
| 177 |
+
datastore = update_datastore(version)
|
| 178 |
hidden_df_elem = get_leaderboard_table(datastore.qa_raw_df, datastore.qa_types, visible=False)
|
| 179 |
+
return hidden_df_elem
|
| 180 |
+
|
| 181 |
+
# def update_qa_version(version):
|
| 182 |
+
# global datastore
|
| 183 |
+
# global ds_dict
|
| 184 |
+
# datastore = ds_dict[version]
|
| 185 |
+
# # domain_elem = get_domain_dropdown(QABenchmarks[datastore.slug])
|
| 186 |
+
# # lang_elem = get_language_dropdown(QABenchmarks[datastore.slug])
|
| 187 |
+
# # model_elem = get_reranking_dropdown(datastore.reranking_models)
|
| 188 |
+
# df_elem = get_leaderboard_table(datastore.qa_fmt_df, datastore.qa_types)
|
| 189 |
+
# hidden_df_elem = get_leaderboard_table(datastore.qa_raw_df, datastore.qa_types, visible=False)
|
| 190 |
+
# return model_elem, df_elem, hidden_df_elem
|
| 191 |
|
| 192 |
|
| 193 |
def update_doc_version(version):
|
|
|
|
| 221 |
# select domain
|
| 222 |
with gr.Row():
|
| 223 |
domains = get_domain_dropdown(QABenchmarks[datastore.slug])
|
| 224 |
+
version.change(
|
| 225 |
+
update_qa_domains,
|
| 226 |
+
version,
|
| 227 |
+
domains
|
| 228 |
+
)
|
| 229 |
# select language
|
| 230 |
with gr.Row():
|
| 231 |
langs = get_language_dropdown(QABenchmarks[datastore.slug])
|
| 232 |
+
version.change(
|
| 233 |
+
update_qa_langs,
|
| 234 |
+
version,
|
| 235 |
+
langs
|
| 236 |
+
)
|
| 237 |
with gr.Column():
|
| 238 |
# select the metric
|
| 239 |
metric = get_metric_dropdown(METRIC_LIST, DEFAULT_METRIC_QA)
|
|
|
|
| 250 |
# select reranking models
|
| 251 |
with gr.Column():
|
| 252 |
models = get_reranking_dropdown(datastore.reranking_models)
|
| 253 |
+
version.change(
|
| 254 |
+
update_qa_models,
|
| 255 |
+
version,
|
| 256 |
+
models
|
| 257 |
+
)
|
| 258 |
# shown_table
|
| 259 |
qa_df_elem_ret_rerank = get_leaderboard_table(datastore.qa_fmt_df, datastore.qa_types)
|
| 260 |
+
version.change(
|
| 261 |
+
update_qa_df_ret_rerank,
|
| 262 |
+
version,
|
| 263 |
+
qa_df_elem_ret_rerank
|
| 264 |
+
)
|
| 265 |
# Dummy leaderboard for handling the case when the user uses backspace key
|
| 266 |
qa_df_elem_ret_rerank_hidden = get_leaderboard_table(
|
| 267 |
datastore.qa_raw_df, datastore.qa_types, visible=False
|
| 268 |
)
|
|
|
|
| 269 |
version.change(
|
| 270 |
+
update_qa_hidden_df_ret_rerank,
|
| 271 |
version,
|
| 272 |
+
qa_df_elem_ret_rerank_hidden
|
| 273 |
)
|
| 274 |
|
| 275 |
set_listeners(
|
|
|
|
| 299 |
search_bar_ret = get_search_bar()
|
| 300 |
with gr.Column(scale=1):
|
| 301 |
models_ret = get_noreranking_dropdown()
|
| 302 |
+
version.change(
|
| 303 |
+
update_qa_models,
|
| 304 |
+
version,
|
| 305 |
+
models_ret
|
| 306 |
+
)
|
| 307 |
+
|
| 308 |
+
_qa_df_ret = datastore.qa_fmt_df[
|
| 309 |
+
datastore.qa_fmt_df[COL_NAME_RERANKING_MODEL] == "NoReranker"]
|
| 310 |
_qa_df_ret = reset_rank(_qa_df_ret)
|
| 311 |
qa_df_elem_ret = get_leaderboard_table(_qa_df_ret, datastore.qa_types)
|
| 312 |
+
version.change(
|
| 313 |
+
update_qa_df_ret,
|
| 314 |
+
version,
|
| 315 |
+
qa_df_elem_ret
|
| 316 |
+
)
|
| 317 |
|
| 318 |
# Dummy leaderboard for handling the case when the user uses backspace key
|
| 319 |
_qa_df_ret_hidden = datastore.qa_raw_df[
|
|
|
|
| 325 |
)
|
| 326 |
|
| 327 |
version.change(
|
| 328 |
+
update_qa_hidden_df_ret,
|
| 329 |
version,
|
| 330 |
+
qa_df_elem_ret_hidden
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 331 |
)
|
| 332 |
|
| 333 |
set_listeners(
|
|
|
|
| 365 |
with gr.Row():
|
| 366 |
with gr.Column(scale=1):
|
| 367 |
qa_models_rerank = get_reranking_dropdown(qa_rerank_models)
|
| 368 |
+
version.change(
|
| 369 |
+
update_qa_models,
|
| 370 |
+
version,
|
| 371 |
+
qa_models_rerank
|
| 372 |
+
)
|
| 373 |
with gr.Column(scale=1):
|
| 374 |
qa_search_bar_rerank = gr.Textbox(show_label=False, visible=False)
|
| 375 |
qa_df_elem_rerank = get_leaderboard_table(_qa_df_rerank, datastore.qa_types)
|
| 376 |
+
version.change(
|
| 377 |
+
update_qa_df_rerank,
|
| 378 |
+
version,
|
| 379 |
+
qa_df_elem_rerank
|
| 380 |
+
)
|
| 381 |
|
| 382 |
_qa_df_rerank_hidden = datastore.qa_raw_df[
|
| 383 |
datastore.qa_raw_df[COL_NAME_RETRIEVAL_MODEL] == BM25_LINK
|
|
|
|
| 386 |
qa_df_elem_rerank_hidden = get_leaderboard_table(
|
| 387 |
_qa_df_rerank_hidden, datastore.qa_types, visible=False
|
| 388 |
)
|
|
|
|
| 389 |
version.change(
|
| 390 |
+
update_qa_hidden_df_rerank,
|
| 391 |
version,
|
| 392 |
+
qa_df_elem_rerank_hidden
|
| 393 |
)
|
| 394 |
|
| 395 |
set_listeners(
|
|
|
|
| 509 |
version.change(
|
| 510 |
update_doc_version,
|
| 511 |
version,
|
| 512 |
+
[
|
| 513 |
+
domains,
|
| 514 |
+
langs,
|
| 515 |
+
models_ret,
|
| 516 |
+
doc_df_elem_ret,
|
| 517 |
+
doc_df_elem_ret_hidden
|
| 518 |
+
],
|
| 519 |
)
|
| 520 |
|
| 521 |
set_listeners(
|