Spaces:
AIR-Bench
/
Running on CPU Upgrade

nan commited on
Commit
0401aeb
1 Parent(s): fe361b7

refactor: refactor the naming

Browse files
Files changed (1) hide show
  1. app.py +94 -138
app.py CHANGED
@@ -170,20 +170,20 @@ with demo:
170
  with gr.Column():
171
  models = get_reranking_dropdown(datastore.reranking_models)
172
  # shown_table
173
- lb_table = get_leaderboard_table(datastore.qa_fmt_df, datastore.qa_types)
174
  # Dummy leaderboard for handling the case when the user uses backspace key
175
- hidden_lb_table = get_leaderboard_table(datastore.qa_raw_df, datastore.qa_types, visible=False)
176
 
177
  version.change(
178
  update_qa_version,
179
  version,
180
- [domains, langs, models, lb_table, hidden_lb_table],
181
  )
182
 
183
  set_listeners(
184
  "qa",
185
- lb_table,
186
- hidden_lb_table,
187
  search_bar,
188
  version,
189
  domains,
@@ -197,136 +197,114 @@ with demo:
197
  metric.change(
198
  update_qa_metric,
199
  [metric, domains, langs, models, search_bar, show_anonymous, show_rev_ts],
200
- lb_table,
201
  queue=True,
202
  )
203
 
204
  with gr.TabItem("Retrieval Only", id=11):
205
  with gr.Row():
206
  with gr.Column(scale=1):
207
- search_bar_retriever = get_search_bar()
208
  with gr.Column(scale=1):
209
- selected_noreranker = get_noreranking_dropdown()
210
 
211
- lb_df_retriever = datastore.qa_fmt_df[
212
- datastore.qa_fmt_df[COL_NAME_RERANKING_MODEL] == "NoReranker"
213
- ]
214
- lb_df_retriever = reset_rank(lb_df_retriever)
215
- lb_table_retriever = get_leaderboard_table(lb_df_retriever, datastore.qa_types)
216
 
217
  # Dummy leaderboard for handling the case when the user uses backspace key
218
- hidden_lb_df_retriever = datastore.qa_raw_df[
219
- datastore.qa_raw_df[COL_NAME_RERANKING_MODEL] == "NoReranker"
220
- ]
221
- hidden_lb_df_retriever = reset_rank(hidden_lb_df_retriever)
222
- hidden_lb_table_retriever = get_leaderboard_table(
223
- hidden_lb_df_retriever, datastore.qa_types, visible=False
224
- )
225
 
226
  version.change(
227
  update_qa_version,
228
- [
229
- version,
230
- ],
231
- [
232
- domains,
233
- langs,
234
- selected_noreranker,
235
- lb_table_retriever,
236
- hidden_lb_table_retriever,
237
- ],
238
  )
239
 
240
  set_listeners(
241
  "qa",
242
- lb_table_retriever,
243
- hidden_lb_table_retriever,
244
- search_bar_retriever,
245
  version,
246
  domains,
247
  langs,
248
- selected_noreranker,
249
  show_anonymous,
250
  show_rev_ts,
251
  )
252
 
253
- # set metric listener
254
  metric.change(
255
  update_qa_metric,
256
  [
257
  metric,
258
  domains,
259
  langs,
260
- selected_noreranker,
261
- search_bar_retriever,
262
  show_anonymous,
263
  show_rev_ts,
264
  ],
265
- lb_table_retriever,
266
  queue=True,
267
  )
 
268
  with gr.TabItem("Reranking Only", id=12):
269
- lb_df_reranker = datastore.qa_fmt_df[
270
  datastore.qa_fmt_df[COL_NAME_RETRIEVAL_MODEL] == BM25_LINK
271
  ]
272
- lb_df_reranker = reset_rank(lb_df_reranker)
273
- reranking_models_reranker = (
274
- lb_df_reranker[COL_NAME_RERANKING_MODEL].apply(remove_html).unique().tolist()
275
  )
276
  with gr.Row():
277
  with gr.Column(scale=1):
278
- selected_rerankings_reranker = get_reranking_dropdown(reranking_models_reranker)
279
  with gr.Column(scale=1):
280
- search_bar_reranker = gr.Textbox(show_label=False, visible=False)
281
- lb_table_reranker = get_leaderboard_table(lb_df_reranker, datastore.qa_types)
282
 
283
- hidden_lb_df_reranker = datastore.qa_raw_df[
284
  datastore.qa_raw_df[COL_NAME_RETRIEVAL_MODEL] == BM25_LINK
285
  ]
286
- hidden_lb_df_reranker = reset_rank(hidden_lb_df_reranker)
287
- hidden_lb_table_reranker = get_leaderboard_table(
288
- hidden_lb_df_reranker, datastore.qa_types, visible=False
289
  )
290
 
291
  version.change(
292
  update_qa_version,
293
- [
294
- version,
295
- ],
296
- [
297
- domains,
298
- langs,
299
- selected_rerankings_reranker,
300
- lb_table_reranker,
301
- hidden_lb_table_reranker,
302
- ],
303
  )
304
 
305
  set_listeners(
306
  "qa",
307
- lb_table_reranker,
308
- hidden_lb_table_reranker,
309
- search_bar_reranker,
310
  version,
311
  domains,
312
  langs,
313
- selected_rerankings_reranker,
314
  show_anonymous,
315
  show_rev_ts,
316
  )
317
- # set metric listener
318
  metric.change(
319
  update_qa_metric,
320
  [
321
  metric,
322
  domains,
323
  langs,
324
- selected_rerankings_reranker,
325
- search_bar_reranker,
326
  show_anonymous,
327
  show_rev_ts,
328
  ],
329
- lb_table_reranker,
330
  queue=True,
331
  )
332
  with gr.TabItem("Long Doc", elem_id="long-doc-benchmark-tab-table", id=1):
@@ -354,33 +332,25 @@ with demo:
354
  with gr.Column():
355
  models = get_reranking_dropdown(datastore.reranking_models)
356
 
357
- lb_table_long_doc = get_leaderboard_table(
358
  datastore.doc_fmt_df, datastore.doc_types
359
  )
360
 
361
  # Dummy leaderboard for handling the case when the user uses backspace key
362
- hidden_lb_table_long_doc = get_leaderboard_table(
363
  datastore.doc_raw_df, datastore.doc_types, visible=False
364
  )
365
 
366
  version.change(
367
  update_doc_version,
368
- [
369
- version,
370
- ],
371
- [
372
- domains,
373
- langs,
374
- models,
375
- lb_table_long_doc,
376
- hidden_lb_table_long_doc,
377
- ],
378
  )
379
 
380
  set_listeners(
381
  "long-doc",
382
- lb_table_long_doc,
383
- hidden_lb_table_long_doc,
384
  search_bar,
385
  version,
386
  domains,
@@ -402,54 +372,47 @@ with demo:
402
  show_anonymous,
403
  show_rev_ts,
404
  ],
405
- lb_table_long_doc,
406
  queue=True,
407
  )
408
  with gr.TabItem("Retrieval Only", id=21):
409
  with gr.Row():
410
  with gr.Column(scale=1):
411
- search_bar_retriever = get_search_bar()
412
  with gr.Column(scale=1):
413
- selected_noreranker = get_noreranking_dropdown()
414
- lb_df_retriever_long_doc = datastore.doc_fmt_df[
 
415
  datastore.doc_fmt_df[COL_NAME_RERANKING_MODEL] == "NoReranker"
416
  ]
417
- lb_df_retriever_long_doc = reset_rank(lb_df_retriever_long_doc)
418
- lb_table_retriever_long_doc = get_leaderboard_table(
419
- lb_df_retriever_long_doc, datastore.doc_types
420
  )
421
 
422
- hidden_lb_df_retriever_long_doc = datastore.doc_raw_df[
423
  datastore.doc_raw_df[COL_NAME_RERANKING_MODEL] == "NoReranker"
424
  ]
425
- hidden_lb_df_retriever_long_doc = reset_rank(hidden_lb_df_retriever_long_doc)
426
- hidden_lb_table_retriever_long_doc = get_leaderboard_table(
427
- hidden_lb_df_retriever_long_doc, datastore.doc_types, visible=False
428
  )
429
 
430
  version.change(
431
  update_doc_version,
432
- [
433
- version,
434
- ],
435
- [
436
- domains,
437
- langs,
438
- selected_noreranker,
439
- lb_table_retriever_long_doc,
440
- hidden_lb_table_retriever_long_doc,
441
- ],
442
  )
443
 
444
  set_listeners(
445
  "long-doc",
446
- lb_table_retriever_long_doc,
447
- hidden_lb_table_retriever_long_doc,
448
- search_bar_retriever,
449
  version,
450
  domains,
451
  langs,
452
- selected_noreranker,
453
  show_anonymous,
454
  show_rev_ts,
455
  )
@@ -460,76 +423,69 @@ with demo:
460
  metric,
461
  domains,
462
  langs,
463
- selected_noreranker,
464
- search_bar_retriever,
465
  show_anonymous,
466
  show_rev_ts,
467
  ],
468
- lb_table_retriever_long_doc,
469
  queue=True,
470
  )
471
  with gr.TabItem("Reranking Only", id=22):
472
- lb_df_reranker_ldoc = datastore.doc_fmt_df[
473
  datastore.doc_fmt_df[COL_NAME_RETRIEVAL_MODEL] == BM25_LINK
474
  ]
475
- lb_df_reranker_ldoc = reset_rank(lb_df_reranker_ldoc)
476
- reranking_models_reranker_ldoc = (
477
- lb_df_reranker_ldoc[COL_NAME_RERANKING_MODEL].apply(remove_html).unique().tolist()
478
  )
479
  with gr.Row():
480
  with gr.Column(scale=1):
481
- selected_rerankings_reranker_ldoc = get_reranking_dropdown(
482
- reranking_models_reranker_ldoc
483
  )
484
  with gr.Column(scale=1):
485
- search_bar_reranker_ldoc = gr.Textbox(show_label=False, visible=False)
486
- lb_table_reranker_ldoc = get_leaderboard_table(lb_df_reranker_ldoc, datastore.doc_types)
487
- hidden_lb_df_reranker_ldoc = datastore.doc_raw_df[
488
  datastore.doc_raw_df[COL_NAME_RETRIEVAL_MODEL] == BM25_LINK
489
  ]
490
- hidden_lb_df_reranker_ldoc = reset_rank(hidden_lb_df_reranker_ldoc)
491
- hidden_lb_table_reranker_ldoc = get_leaderboard_table(
492
- hidden_lb_df_reranker_ldoc, datastore.doc_types, visible=False
493
  )
494
 
495
  version.change(
496
  update_doc_version,
497
- [
498
- version,
499
- ],
500
- [
501
- domains,
502
- langs,
503
- selected_rerankings_reranker_ldoc,
504
- lb_table_reranker_ldoc,
505
- hidden_lb_table_reranker_ldoc,
506
- ],
507
  )
508
 
509
  set_listeners(
510
  "long-doc",
511
- lb_table_reranker_ldoc,
512
- hidden_lb_table_reranker_ldoc,
513
- search_bar_reranker_ldoc,
514
  version,
515
  domains,
516
  langs,
517
- selected_rerankings_reranker_ldoc,
518
  show_anonymous,
519
  show_rev_ts,
520
  )
 
521
  metric.change(
522
  update_doc_metric,
523
  [
524
  metric,
525
  domains,
526
  langs,
527
- selected_rerankings_reranker_ldoc,
528
- search_bar_reranker_ldoc,
529
  show_anonymous,
530
  show_rev_ts,
531
  ],
532
- lb_table_reranker_ldoc,
533
  queue=True,
534
  )
535
 
 
170
  with gr.Column():
171
  models = get_reranking_dropdown(datastore.reranking_models)
172
  # shown_table
173
+ qa_df_elem_ret_rerank = get_leaderboard_table(datastore.qa_fmt_df, datastore.qa_types)
174
  # Dummy leaderboard for handling the case when the user uses backspace key
175
+ qa_df_elem_ret_rerank_hidden = get_leaderboard_table(datastore.qa_raw_df, datastore.qa_types, visible=False)
176
 
177
  version.change(
178
  update_qa_version,
179
  version,
180
+ [domains, langs, models, qa_df_elem_ret_rerank, qa_df_elem_ret_rerank_hidden],
181
  )
182
 
183
  set_listeners(
184
  "qa",
185
+ qa_df_elem_ret_rerank,
186
+ qa_df_elem_ret_rerank_hidden,
187
  search_bar,
188
  version,
189
  domains,
 
197
  metric.change(
198
  update_qa_metric,
199
  [metric, domains, langs, models, search_bar, show_anonymous, show_rev_ts],
200
+ qa_df_elem_ret_rerank,
201
  queue=True,
202
  )
203
 
204
  with gr.TabItem("Retrieval Only", id=11):
205
  with gr.Row():
206
  with gr.Column(scale=1):
207
+ search_bar_ret = get_search_bar()
208
  with gr.Column(scale=1):
209
+ models_ret = get_noreranking_dropdown()
210
 
211
+ _qa_df_ret = datastore.qa_fmt_df[datastore.qa_fmt_df[COL_NAME_RERANKING_MODEL] == "NoReranker"]
212
+ _qa_df_ret = reset_rank(_qa_df_ret)
213
+ qa_df_elem_ret = get_leaderboard_table(_qa_df_ret, datastore.qa_types)
 
 
214
 
215
  # Dummy leaderboard for handling the case when the user uses backspace key
216
+ _qa_df_ret_hidden = datastore.qa_raw_df[datastore.qa_raw_df[COL_NAME_RERANKING_MODEL] == "NoReranker"]
217
+ _qa_df_ret_hidden = reset_rank(_qa_df_ret_hidden)
218
+ qa_df_elem_ret_hidden = get_leaderboard_table(_qa_df_ret_hidden, datastore.qa_types, visible=False)
 
 
 
 
219
 
220
  version.change(
221
  update_qa_version,
222
+ version,
223
+ [domains, langs, models_ret, qa_df_elem_ret, qa_df_elem_ret_hidden, ],
 
 
 
 
 
 
 
 
224
  )
225
 
226
  set_listeners(
227
  "qa",
228
+ qa_df_elem_ret,
229
+ qa_df_elem_ret_hidden,
230
+ search_bar_ret,
231
  version,
232
  domains,
233
  langs,
234
+ models_ret,
235
  show_anonymous,
236
  show_rev_ts,
237
  )
238
 
 
239
  metric.change(
240
  update_qa_metric,
241
  [
242
  metric,
243
  domains,
244
  langs,
245
+ models_ret,
246
+ search_bar_ret,
247
  show_anonymous,
248
  show_rev_ts,
249
  ],
250
+ qa_df_elem_ret,
251
  queue=True,
252
  )
253
+
254
  with gr.TabItem("Reranking Only", id=12):
255
+ _qa_df_rerank = datastore.qa_fmt_df[
256
  datastore.qa_fmt_df[COL_NAME_RETRIEVAL_MODEL] == BM25_LINK
257
  ]
258
+ _qa_df_rerank = reset_rank(_qa_df_rerank)
259
+ qa_rerank_models = (
260
+ _qa_df_rerank[COL_NAME_RERANKING_MODEL].apply(remove_html).unique().tolist()
261
  )
262
  with gr.Row():
263
  with gr.Column(scale=1):
264
+ qa_models_rerank = get_reranking_dropdown(qa_rerank_models)
265
  with gr.Column(scale=1):
266
+ qa_search_bar_rerank = gr.Textbox(show_label=False, visible=False)
267
+ qa_df_elem_rerank = get_leaderboard_table(_qa_df_rerank, datastore.qa_types)
268
 
269
+ _qa_df_rerank_hidden = datastore.qa_raw_df[
270
  datastore.qa_raw_df[COL_NAME_RETRIEVAL_MODEL] == BM25_LINK
271
  ]
272
+ _qa_df_rerank_hidden = reset_rank(_qa_df_rerank_hidden)
273
+ qa_df_elem_rerank_hidden = get_leaderboard_table(
274
+ _qa_df_rerank_hidden, datastore.qa_types, visible=False
275
  )
276
 
277
  version.change(
278
  update_qa_version,
279
+ version,
280
+ [domains, langs, qa_models_rerank, qa_df_elem_rerank, qa_df_elem_rerank_hidden],
 
 
 
 
 
 
 
 
281
  )
282
 
283
  set_listeners(
284
  "qa",
285
+ qa_df_elem_rerank,
286
+ qa_df_elem_rerank_hidden,
287
+ qa_search_bar_rerank,
288
  version,
289
  domains,
290
  langs,
291
+ qa_models_rerank,
292
  show_anonymous,
293
  show_rev_ts,
294
  )
295
+
296
  metric.change(
297
  update_qa_metric,
298
  [
299
  metric,
300
  domains,
301
  langs,
302
+ qa_models_rerank,
303
+ qa_search_bar_rerank,
304
  show_anonymous,
305
  show_rev_ts,
306
  ],
307
+ qa_df_elem_rerank,
308
  queue=True,
309
  )
310
  with gr.TabItem("Long Doc", elem_id="long-doc-benchmark-tab-table", id=1):
 
332
  with gr.Column():
333
  models = get_reranking_dropdown(datastore.reranking_models)
334
 
335
+ doc_df_elem_ret_rerank = get_leaderboard_table(
336
  datastore.doc_fmt_df, datastore.doc_types
337
  )
338
 
339
  # Dummy leaderboard for handling the case when the user uses backspace key
340
+ doc_df_elem_ret_rerank_hidden = get_leaderboard_table(
341
  datastore.doc_raw_df, datastore.doc_types, visible=False
342
  )
343
 
344
  version.change(
345
  update_doc_version,
346
+ version,
347
+ [domains, langs, models, doc_df_elem_ret_rerank, doc_df_elem_ret_rerank_hidden],
 
 
 
 
 
 
 
 
348
  )
349
 
350
  set_listeners(
351
  "long-doc",
352
+ doc_df_elem_ret_rerank,
353
+ doc_df_elem_ret_rerank_hidden,
354
  search_bar,
355
  version,
356
  domains,
 
372
  show_anonymous,
373
  show_rev_ts,
374
  ],
375
+ doc_df_elem_ret_rerank,
376
  queue=True,
377
  )
378
  with gr.TabItem("Retrieval Only", id=21):
379
  with gr.Row():
380
  with gr.Column(scale=1):
381
+ search_bar_ret = get_search_bar()
382
  with gr.Column(scale=1):
383
+ models_ret = get_noreranking_dropdown()
384
+
385
+ _doc_df_ret = datastore.doc_fmt_df[
386
  datastore.doc_fmt_df[COL_NAME_RERANKING_MODEL] == "NoReranker"
387
  ]
388
+ _doc_df_ret = reset_rank(_doc_df_ret)
389
+ doc_df_elem_ret = get_leaderboard_table(
390
+ _doc_df_ret, datastore.doc_types
391
  )
392
 
393
+ _doc_df_ret_hidden = datastore.doc_raw_df[
394
  datastore.doc_raw_df[COL_NAME_RERANKING_MODEL] == "NoReranker"
395
  ]
396
+ _doc_df_ret_hidden = reset_rank(_doc_df_ret_hidden)
397
+ doc_df_elem_ret_hidden = get_leaderboard_table(
398
+ _doc_df_ret_hidden, datastore.doc_types, visible=False
399
  )
400
 
401
  version.change(
402
  update_doc_version,
403
+ version,
404
+ [domains, langs, models_ret, doc_df_elem_ret, doc_df_elem_ret_hidden],
 
 
 
 
 
 
 
 
405
  )
406
 
407
  set_listeners(
408
  "long-doc",
409
+ doc_df_elem_ret,
410
+ doc_df_elem_ret_hidden,
411
+ search_bar_ret,
412
  version,
413
  domains,
414
  langs,
415
+ models_ret,
416
  show_anonymous,
417
  show_rev_ts,
418
  )
 
423
  metric,
424
  domains,
425
  langs,
426
+ models_ret,
427
+ search_bar_ret,
428
  show_anonymous,
429
  show_rev_ts,
430
  ],
431
+ doc_df_elem_ret,
432
  queue=True,
433
  )
434
  with gr.TabItem("Reranking Only", id=22):
435
+ _doc_df_rerank = datastore.doc_fmt_df[
436
  datastore.doc_fmt_df[COL_NAME_RETRIEVAL_MODEL] == BM25_LINK
437
  ]
438
+ _doc_df_rerank = reset_rank(_doc_df_rerank)
439
+ doc_rerank_models = (
440
+ _doc_df_rerank[COL_NAME_RERANKING_MODEL].apply(remove_html).unique().tolist()
441
  )
442
  with gr.Row():
443
  with gr.Column(scale=1):
444
+ doc_models_rerank = get_reranking_dropdown(
445
+ doc_rerank_models
446
  )
447
  with gr.Column(scale=1):
448
+ doc_search_bar_rerank = gr.Textbox(show_label=False, visible=False)
449
+ doc_df_elem_rerank = get_leaderboard_table(_doc_df_rerank, datastore.doc_types)
450
+ _doc_df_rerank_hidden = datastore.doc_raw_df[
451
  datastore.doc_raw_df[COL_NAME_RETRIEVAL_MODEL] == BM25_LINK
452
  ]
453
+ _doc_df_rerank_hidden = reset_rank(_doc_df_rerank_hidden)
454
+ doc_df_elem_rerank_hidden = get_leaderboard_table(
455
+ _doc_df_rerank_hidden, datastore.doc_types, visible=False
456
  )
457
 
458
  version.change(
459
  update_doc_version,
460
+ version,
461
+ [domains, langs, doc_models_rerank, doc_df_elem_rerank, doc_df_elem_rerank_hidden],
 
 
 
 
 
 
 
 
462
  )
463
 
464
  set_listeners(
465
  "long-doc",
466
+ doc_df_elem_rerank,
467
+ doc_df_elem_rerank_hidden,
468
+ doc_search_bar_rerank,
469
  version,
470
  domains,
471
  langs,
472
+ doc_models_rerank,
473
  show_anonymous,
474
  show_rev_ts,
475
  )
476
+
477
  metric.change(
478
  update_doc_metric,
479
  [
480
  metric,
481
  domains,
482
  langs,
483
+ doc_models_rerank,
484
+ doc_search_bar_rerank,
485
  show_anonymous,
486
  show_rev_ts,
487
  ],
488
+ doc_df_elem_rerank,
489
  queue=True,
490
  )
491