Spaces:
AIR-Bench
/
Running on CPU Upgrade

nan commited on
Commit
f70b03c
1 Parent(s): bf586e3

fix: fix the bug in loading selected reranker

Browse files
Files changed (1) hide show
  1. app.py +79 -29
app.py CHANGED
@@ -82,6 +82,7 @@ def update_metric_long_doc(
82
 
83
 
84
  def update_datastore(version):
 
85
  global datastore
86
  global data
87
  datastore = data[version]
@@ -89,12 +90,26 @@ def update_datastore(version):
89
  selected_langs = get_language_dropdown(QABenchmarks[datastore.slug])
90
  selected_rerankings = get_reranking_dropdown(datastore.reranking_models)
91
  leaderboard_table = get_leaderboard_table(
92
- datastore.raw_df_qa, datastore.types_qa)
93
- hidden_leaderboard_table_for_search = get_leaderboard_table(
94
  datastore.raw_df_qa, datastore.types_qa, visible=False)
95
- return selected_domains, selected_langs, selected_rerankings, leaderboard_table, hidden_leaderboard_table_for_search
96
- # DOMAIN_COLS_LONG_DOC = list(frozenset([c.value.domain for c in list(LongDocBenchmarks)]))
97
- # LANG_COLS_LONG_DOC = list(frozenset([c.value.lang for c in list(LongDocBenchmarks)]))
 
 
 
 
 
 
 
 
 
 
 
 
 
 
98
 
99
  demo = gr.Blocks(css=custom_css)
100
 
@@ -183,7 +198,8 @@ with demo:
183
 
184
  lb_df_retriever = datastore.leaderboard_df_qa[datastore.leaderboard_df_qa[COL_NAME_RERANKING_MODEL] == "NoReranker"]
185
  lb_df_retriever = reset_rank(lb_df_retriever)
186
- lb_table_retriever = get_leaderboard_table(lb_df_retriever, datastore.types_qa)
 
187
 
188
  # Dummy leaderboard for handling the case when the user uses backspace key
189
  hidden_lb_df_retriever = datastore.raw_df_qa[datastore.raw_df_qa[COL_NAME_RERANKING_MODEL] == "NoReranker"]
@@ -193,7 +209,13 @@ with demo:
193
  selected_version.change(
194
  update_datastore,
195
  [selected_version,],
196
- [selected_domains, selected_langs, selected_rerankings, lb_table_retriever, hidden_lb_table_retriever]
 
 
 
 
 
 
197
  )
198
 
199
  set_listeners(
@@ -233,18 +255,26 @@ with demo:
233
  selected_rerankings_reranker = get_reranking_dropdown(reranking_models_reranker)
234
  with gr.Column(scale=1):
235
  search_bar_reranker = gr.Textbox(show_label=False, visible=False)
236
- lb_table_reranker = get_leaderboard_table(lb_df_reranker, datastore.types_qa)
 
237
 
238
  hidden_lb_df_reranker = datastore.raw_df_qa[datastore.raw_df_qa[COL_NAME_RETRIEVAL_MODEL] == BM25_LINK]
239
  hidden_lb_df_reranker = reset_rank(hidden_lb_df_reranker)
240
  hidden_lb_table_reranker = get_leaderboard_table(
241
- hidden_lb_df_reranker, datastore.types_qa, visible=False
 
242
  )
243
 
244
  selected_version.change(
245
  update_datastore,
246
  [selected_version,],
247
- [selected_domains, selected_langs, selected_rerankings_reranker, lb_table_reranker, hidden_lb_table_reranker]
 
 
 
 
 
 
248
  )
249
 
250
  set_listeners(
@@ -300,25 +330,31 @@ with demo:
300
  with gr.Column():
301
  selected_rerankings = get_reranking_dropdown(datastore.reranking_models)
302
 
303
- lb_table = get_leaderboard_table(
304
  datastore.leaderboard_df_long_doc, datastore.types_long_doc
305
  )
306
 
307
  # Dummy leaderboard for handling the case when the user uses backspace key
308
- hidden_lb_table = get_leaderboard_table(
309
  datastore.raw_df_long_doc, datastore.types_long_doc, visible=False
310
  )
311
 
312
  selected_version.change(
313
- update_datastore,
314
  [selected_version,],
315
- [selected_domains, selected_langs, selected_rerankings, lb_table, hidden_lb_table]
 
 
 
 
 
 
316
  )
317
 
318
  set_listeners(
319
  "long-doc",
320
- lb_table,
321
- hidden_lb_table,
322
  search_bar,
323
  selected_version,
324
  selected_domains,
@@ -340,7 +376,7 @@ with demo:
340
  show_anonymous,
341
  show_revision_and_timestamp
342
  ],
343
- lb_table,
344
  queue=True
345
  )
346
  """
@@ -350,18 +386,31 @@ with demo:
350
  search_bar_retriever = get_search_bar()
351
  with gr.Column(scale=1):
352
  selected_noreranker = get_noreranking_dropdown()
353
- lb_df_retriever_long_doc = data["AIR-Bench_24.04"].leaderboard_df_long_doc[
354
- data["AIR-Bench_24.04"].leaderboard_df_long_doc[COL_NAME_RERANKING_MODEL] == "NoReranker"
355
  ]
356
  lb_df_retriever_long_doc = reset_rank(lb_df_retriever_long_doc)
357
- hidden_lb_db_retriever_long_doc = data["AIR-Bench_24.04"].raw_df_long_doc[
358
- data["AIR-Bench_24.04"].raw_df_long_doc[COL_NAME_RERANKING_MODEL] == "NoReranker"
 
 
 
359
  ]
360
  hidden_lb_db_retriever_long_doc = reset_rank(hidden_lb_db_retriever_long_doc)
361
- lb_table_retriever_long_doc = get_leaderboard_table(
362
- lb_df_retriever_long_doc, data["AIR-Bench_24.04"].types_long_doc)
363
  hidden_lb_table_retriever_long_doc = get_leaderboard_table(
364
- hidden_lb_db_retriever_long_doc, data["AIR-Bench_24.04"].types_long_doc, visible=False
 
 
 
 
 
 
 
 
 
 
 
 
365
  )
366
 
367
  set_listeners(
@@ -369,6 +418,7 @@ with demo:
369
  lb_table_retriever_long_doc,
370
  hidden_lb_table_retriever_long_doc,
371
  search_bar_retriever,
 
372
  selected_domains,
373
  selected_langs,
374
  selected_noreranker,
@@ -391,8 +441,8 @@ with demo:
391
  queue=True
392
  )
393
  with gr.TabItem("Reranking Only", id=22):
394
- lb_df_reranker_ldoc = data["AIR-Bench_24.04"].leaderboard_df_long_doc[
395
- data["AIR-Bench_24.04"].leaderboard_df_long_doc[COL_NAME_RETRIEVAL_MODEL] == BM25_LINK
396
  ]
397
  lb_df_reranker_ldoc = reset_rank(lb_df_reranker_ldoc)
398
  reranking_models_reranker_ldoc = lb_df_reranker_ldoc[COL_NAME_RERANKING_MODEL].apply(remove_html).unique().tolist()
@@ -401,11 +451,11 @@ with demo:
401
  selected_rerankings_reranker_ldoc = get_reranking_dropdown(reranking_models_reranker_ldoc)
402
  with gr.Column(scale=1):
403
  search_bar_reranker_ldoc = gr.Textbox(show_label=False, visible=False)
404
- lb_table_reranker_ldoc = get_leaderboard_table(lb_df_reranker_ldoc, data["AIR-Bench_24.04"].types_long_doc)
405
- hidden_lb_df_reranker_ldoc = data["AIR-Bench_24.04"].raw_df_long_doc[data["AIR-Bench_24.04"].raw_df_long_doc[COL_NAME_RETRIEVAL_MODEL] == BM25_LINK]
406
  hidden_lb_df_reranker_ldoc = reset_rank(hidden_lb_df_reranker_ldoc)
407
  hidden_lb_table_reranker_ldoc = get_leaderboard_table(
408
- hidden_lb_df_reranker_ldoc, data["AIR-Bench_24.04"].types_long_doc, visible=False
409
  )
410
 
411
  set_listeners(
 
82
 
83
 
84
  def update_datastore(version):
85
+ print("triggered update_datastore")
86
  global datastore
87
  global data
88
  datastore = data[version]
 
90
  selected_langs = get_language_dropdown(QABenchmarks[datastore.slug])
91
  selected_rerankings = get_reranking_dropdown(datastore.reranking_models)
92
  leaderboard_table = get_leaderboard_table(
93
+ datastore.leaderboard_df_qa, datastore.types_qa)
94
+ hidden_leaderboard_table = get_leaderboard_table(
95
  datastore.raw_df_qa, datastore.types_qa, visible=False)
96
+ return selected_domains, selected_langs, selected_rerankings, leaderboard_table, hidden_leaderboard_table
97
+
98
+
99
+ def update_datastore_long_doc(version):
100
+ global datastore
101
+ global data
102
+ print("triggered update_datastore_long_doc")
103
+ datastore = data[version]
104
+ selected_domains = get_domain_dropdown(LongDocBenchmarks[datastore.slug])
105
+ selected_langs = get_language_dropdown(LongDocBenchmarks[datastore.slug])
106
+ selected_rerankings = get_reranking_dropdown(datastore.reranking_models)
107
+ leaderboard_table = get_leaderboard_table(
108
+ datastore.leaderboard_df_long_doc, datastore.types_long_doc)
109
+ hidden_leaderboard_table = get_leaderboard_table(
110
+ datastore.raw_df_long_doc, datastore.types_long_doc, visible=False)
111
+ return selected_domains, selected_langs, selected_rerankings, leaderboard_table, hidden_leaderboard_table
112
+
113
 
114
  demo = gr.Blocks(css=custom_css)
115
 
 
198
 
199
  lb_df_retriever = datastore.leaderboard_df_qa[datastore.leaderboard_df_qa[COL_NAME_RERANKING_MODEL] == "NoReranker"]
200
  lb_df_retriever = reset_rank(lb_df_retriever)
201
+ lb_table_retriever = get_leaderboard_table(
202
+ lb_df_retriever, datastore.types_qa)
203
 
204
  # Dummy leaderboard for handling the case when the user uses backspace key
205
  hidden_lb_df_retriever = datastore.raw_df_qa[datastore.raw_df_qa[COL_NAME_RERANKING_MODEL] == "NoReranker"]
 
209
  selected_version.change(
210
  update_datastore,
211
  [selected_version,],
212
+ [
213
+ selected_domains,
214
+ selected_langs,
215
+ selected_noreranker,
216
+ lb_table_retriever,
217
+ hidden_lb_table_retriever
218
+ ]
219
  )
220
 
221
  set_listeners(
 
255
  selected_rerankings_reranker = get_reranking_dropdown(reranking_models_reranker)
256
  with gr.Column(scale=1):
257
  search_bar_reranker = gr.Textbox(show_label=False, visible=False)
258
+ lb_table_reranker = get_leaderboard_table(
259
+ lb_df_reranker, datastore.types_qa)
260
 
261
  hidden_lb_df_reranker = datastore.raw_df_qa[datastore.raw_df_qa[COL_NAME_RETRIEVAL_MODEL] == BM25_LINK]
262
  hidden_lb_df_reranker = reset_rank(hidden_lb_df_reranker)
263
  hidden_lb_table_reranker = get_leaderboard_table(
264
+ hidden_lb_df_reranker,
265
+ datastore.types_qa, visible=False
266
  )
267
 
268
  selected_version.change(
269
  update_datastore,
270
  [selected_version,],
271
+ [
272
+ selected_domains,
273
+ selected_langs,
274
+ selected_rerankings_reranker,
275
+ lb_table_reranker,
276
+ hidden_lb_table_reranker
277
+ ]
278
  )
279
 
280
  set_listeners(
 
330
  with gr.Column():
331
  selected_rerankings = get_reranking_dropdown(datastore.reranking_models)
332
 
333
+ lb_table_long_doc = get_leaderboard_table(
334
  datastore.leaderboard_df_long_doc, datastore.types_long_doc
335
  )
336
 
337
  # Dummy leaderboard for handling the case when the user uses backspace key
338
+ hidden_lb_table_long_doc = get_leaderboard_table(
339
  datastore.raw_df_long_doc, datastore.types_long_doc, visible=False
340
  )
341
 
342
  selected_version.change(
343
+ update_datastore_long_doc,
344
  [selected_version,],
345
+ [
346
+ selected_domains,
347
+ selected_langs,
348
+ selected_rerankings,
349
+ lb_table_long_doc,
350
+ hidden_lb_table_long_doc
351
+ ]
352
  )
353
 
354
  set_listeners(
355
  "long-doc",
356
+ lb_table_long_doc,
357
+ hidden_lb_table_long_doc,
358
  search_bar,
359
  selected_version,
360
  selected_domains,
 
376
  show_anonymous,
377
  show_revision_and_timestamp
378
  ],
379
+ lb_table_long_doc,
380
  queue=True
381
  )
382
  """
 
386
  search_bar_retriever = get_search_bar()
387
  with gr.Column(scale=1):
388
  selected_noreranker = get_noreranking_dropdown()
389
+ lb_df_retriever_long_doc = datastore.leaderboard_df_long_doc[
390
+ datastore.leaderboard_df_long_doc[COL_NAME_RERANKING_MODEL] == "NoReranker"
391
  ]
392
  lb_df_retriever_long_doc = reset_rank(lb_df_retriever_long_doc)
393
+ lb_table_retriever_long_doc = get_leaderboard_table(
394
+ lb_df_retriever_long_doc, datastore.types_long_doc)
395
+
396
+ hidden_lb_db_retriever_long_doc = datastore.raw_df_long_doc[
397
+ datastore.raw_df_long_doc[COL_NAME_RERANKING_MODEL] == "NoReranker"
398
  ]
399
  hidden_lb_db_retriever_long_doc = reset_rank(hidden_lb_db_retriever_long_doc)
 
 
400
  hidden_lb_table_retriever_long_doc = get_leaderboard_table(
401
+ hidden_lb_db_retriever_long_doc, datastore.types_long_doc, visible=False
402
+ )
403
+
404
+ selected_version.change(
405
+ update_datastore,
406
+ [selected_version,],
407
+ [
408
+ selected_domains,
409
+ selected_langs,
410
+ selected_rerankings,
411
+ lb_table_retriever_long_doc,
412
+ hidden_lb_table_retriever_long_doc
413
+ ]
414
  )
415
 
416
  set_listeners(
 
418
  lb_table_retriever_long_doc,
419
  hidden_lb_table_retriever_long_doc,
420
  search_bar_retriever,
421
+ selected_version,
422
  selected_domains,
423
  selected_langs,
424
  selected_noreranker,
 
441
  queue=True
442
  )
443
  with gr.TabItem("Reranking Only", id=22):
444
+ lb_df_reranker_ldoc = datastore.leaderboard_df_long_doc[
445
+ datastore.leaderboard_df_long_doc[COL_NAME_RETRIEVAL_MODEL] == BM25_LINK
446
  ]
447
  lb_df_reranker_ldoc = reset_rank(lb_df_reranker_ldoc)
448
  reranking_models_reranker_ldoc = lb_df_reranker_ldoc[COL_NAME_RERANKING_MODEL].apply(remove_html).unique().tolist()
 
451
  selected_rerankings_reranker_ldoc = get_reranking_dropdown(reranking_models_reranker_ldoc)
452
  with gr.Column(scale=1):
453
  search_bar_reranker_ldoc = gr.Textbox(show_label=False, visible=False)
454
+ lb_table_reranker_ldoc = get_leaderboard_table(lb_df_reranker_ldoc, datastore.types_long_doc)
455
+ hidden_lb_df_reranker_ldoc = datastore.raw_df_long_doc[datastore.raw_df_long_doc[COL_NAME_RETRIEVAL_MODEL] == BM25_LINK]
456
  hidden_lb_df_reranker_ldoc = reset_rank(hidden_lb_df_reranker_ldoc)
457
  hidden_lb_table_reranker_ldoc = get_leaderboard_table(
458
+ hidden_lb_df_reranker_ldoc, datastore.types_long_doc, visible=False
459
  )
460
 
461
  set_listeners(