kon172verma commited on
Commit
d51152f
·
verified ·
1 Parent(s): 6174492

Add LoRA+ test report: smollm2-360m_A_1k [20260826_203741]

Browse files
reports/v2.0/loraplus/reports_test/smollm2-360m_A_1k_test_20260826_203741.json ADDED
@@ -0,0 +1,1102 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "model_key": "smollm2-360m",
3
+ "model_id": "HuggingFaceTB/SmolLM2-360M-Instruct",
4
+ "lora_config": "A",
5
+ "dataset_size": "1k",
6
+ "technique": "LoRA+",
7
+ "split": "test",
8
+ "timestamp": "20260826_203741",
9
+ "prompt_format": "tool_id_v1",
10
+ "tool_id_scheme": "abcdefghijklmnopqrstuvwxyzABCDEFGHIJKLMNOPQRSTUVWXYZ",
11
+ "no_tool_id": "-",
12
+ "n_examples": 100,
13
+ "n_correct": 67,
14
+ "accuracy": 0.67,
15
+ "anchor_accuracy": null,
16
+ "avg_latency_ms": 105.21,
17
+ "p50_latency_ms": 99.98,
18
+ "p95_latency_ms": 102.94,
19
+ "avg_tokens_per_sec": 19.81,
20
+ "peak_memory_mb": 758.3,
21
+ "hf_repo": "kon172verma/intent-classifier-experiments",
22
+ "hf_subfolder": "v2.0/smollm2-360m_LoRA+_A_1k_20260826-164600",
23
+ "hf_report_path": "reports/v2.0/loraplus/reports_test/smollm2-360m_A_1k_test_20260826_203741.json",
24
+ "per_tool_metrics": {
25
+ "battery_status_monitor": {
26
+ "precision": 0.25,
27
+ "recall": 0.5,
28
+ "f1": 0.3333,
29
+ "support": 2
30
+ },
31
+ "call_handler": {
32
+ "precision": 0.6,
33
+ "recall": 0.6,
34
+ "f1": 0.6,
35
+ "support": 5
36
+ },
37
+ "climate_control": {
38
+ "precision": 1.0,
39
+ "recall": 1.0,
40
+ "f1": 1.0,
41
+ "support": 1
42
+ },
43
+ "corp-cloud-sync": {
44
+ "precision": 0.8571,
45
+ "recall": 0.8571,
46
+ "f1": 0.8571,
47
+ "support": 7
48
+ },
49
+ "corp-contact-sync": {
50
+ "precision": 1.0,
51
+ "recall": 0.4286,
52
+ "f1": 0.6,
53
+ "support": 7
54
+ },
55
+ "corp-fleet-manager": {
56
+ "precision": 0.5,
57
+ "recall": 1.0,
58
+ "f1": 0.6667,
59
+ "support": 1
60
+ },
61
+ "corp-kariya-weather-app": {
62
+ "precision": 0.0,
63
+ "recall": 0.0,
64
+ "f1": 0.0,
65
+ "support": 0
66
+ },
67
+ "corp-scheduler-app": {
68
+ "precision": 0.5,
69
+ "recall": 0.3333,
70
+ "f1": 0.4,
71
+ "support": 3
72
+ },
73
+ "ev_charging_scheduler": {
74
+ "precision": 0.5,
75
+ "recall": 0.5,
76
+ "f1": 0.5,
77
+ "support": 2
78
+ },
79
+ "face_emotion": {
80
+ "precision": 0.6667,
81
+ "recall": 0.5,
82
+ "f1": 0.5714,
83
+ "support": 4
84
+ },
85
+ "fuel_station_finder": {
86
+ "precision": 0.75,
87
+ "recall": 0.75,
88
+ "f1": 0.75,
89
+ "support": 4
90
+ },
91
+ "health_monitor": {
92
+ "precision": 1.0,
93
+ "recall": 0.6,
94
+ "f1": 0.75,
95
+ "support": 5
96
+ },
97
+ "in_car_analytics": {
98
+ "precision": 1.0,
99
+ "recall": 0.75,
100
+ "f1": 0.8571,
101
+ "support": 4
102
+ },
103
+ "insurance_claims": {
104
+ "precision": 0.0,
105
+ "recall": 0.0,
106
+ "f1": 0.0,
107
+ "support": 0
108
+ },
109
+ "media_player_ctrl": {
110
+ "precision": 1.0,
111
+ "recall": 1.0,
112
+ "f1": 1.0,
113
+ "support": 1
114
+ },
115
+ "nav_route_planner": {
116
+ "precision": 1.0,
117
+ "recall": 0.25,
118
+ "f1": 0.4,
119
+ "support": 4
120
+ },
121
+ "news_briefing": {
122
+ "precision": 0.25,
123
+ "recall": 0.5,
124
+ "f1": 0.3333,
125
+ "support": 2
126
+ },
127
+ "none": {
128
+ "precision": 0.9524,
129
+ "recall": 1.0,
130
+ "f1": 0.9756,
131
+ "support": 20
132
+ },
133
+ "package_tracker": {
134
+ "precision": 0.6667,
135
+ "recall": 0.5,
136
+ "f1": 0.5714,
137
+ "support": 4
138
+ },
139
+ "parking_locator": {
140
+ "precision": 1.0,
141
+ "recall": 0.6667,
142
+ "f1": 0.8,
143
+ "support": 6
144
+ },
145
+ "poi_search": {
146
+ "precision": 0.3333,
147
+ "recall": 1.0,
148
+ "f1": 0.5,
149
+ "support": 1
150
+ },
151
+ "remote_vehicle_lock": {
152
+ "precision": 0.0,
153
+ "recall": 0.0,
154
+ "f1": 0.0,
155
+ "support": 2
156
+ },
157
+ "roadside_assistance": {
158
+ "precision": 0.0,
159
+ "recall": 0.0,
160
+ "f1": 0.0,
161
+ "support": 1
162
+ },
163
+ "safety_proposal": {
164
+ "precision": 0.4286,
165
+ "recall": 1.0,
166
+ "f1": 0.6,
167
+ "support": 3
168
+ },
169
+ "sms_messenger": {
170
+ "precision": 0.5,
171
+ "recall": 0.5,
172
+ "f1": 0.5,
173
+ "support": 2
174
+ },
175
+ "toll_payment_service": {
176
+ "precision": 0.0,
177
+ "recall": 0.0,
178
+ "f1": 0.0,
179
+ "support": 1
180
+ },
181
+ "traffic_alert_service": {
182
+ "precision": 0.6,
183
+ "recall": 0.6,
184
+ "f1": 0.6,
185
+ "support": 5
186
+ },
187
+ "trip_logger": {
188
+ "precision": 1.0,
189
+ "recall": 0.5,
190
+ "f1": 0.6667,
191
+ "support": 2
192
+ },
193
+ "vehicle_diagnostics": {
194
+ "precision": 0.25,
195
+ "recall": 1.0,
196
+ "f1": 0.4,
197
+ "support": 1
198
+ }
199
+ },
200
+ "results": [
201
+ {
202
+ "index": 1,
203
+ "answer": "call_handler",
204
+ "answer_id": "d",
205
+ "prediction": "d",
206
+ "prediction_tool": "call_handler",
207
+ "correct": true,
208
+ "latency_s": 0.6073
209
+ },
210
+ {
211
+ "index": 2,
212
+ "answer": "traffic_alert_service",
213
+ "answer_id": "l",
214
+ "prediction": "e",
215
+ "prediction_tool": "call_handler",
216
+ "correct": false,
217
+ "latency_s": 0.1004
218
+ },
219
+ {
220
+ "index": 3,
221
+ "answer": "health_monitor",
222
+ "answer_id": "b",
223
+ "prediction": "b",
224
+ "prediction_tool": "health_monitor",
225
+ "correct": true,
226
+ "latency_s": 0.1
227
+ },
228
+ {
229
+ "index": 4,
230
+ "answer": "corp-cloud-sync",
231
+ "answer_id": "d",
232
+ "prediction": "d",
233
+ "prediction_tool": "corp-cloud-sync",
234
+ "correct": true,
235
+ "latency_s": 0.0993
236
+ },
237
+ {
238
+ "index": 5,
239
+ "answer": "none",
240
+ "answer_id": "-",
241
+ "prediction": "-",
242
+ "prediction_tool": "none",
243
+ "correct": true,
244
+ "latency_s": 0.1018
245
+ },
246
+ {
247
+ "index": 6,
248
+ "answer": "corp-contact-sync",
249
+ "answer_id": "g",
250
+ "prediction": "-",
251
+ "prediction_tool": "none",
252
+ "correct": false,
253
+ "latency_s": 0.0999
254
+ },
255
+ {
256
+ "index": 7,
257
+ "answer": "parking_locator",
258
+ "answer_id": "o",
259
+ "prediction": "o",
260
+ "prediction_tool": "parking_locator",
261
+ "correct": true,
262
+ "latency_s": 0.1004
263
+ },
264
+ {
265
+ "index": 8,
266
+ "answer": "fuel_station_finder",
267
+ "answer_id": "b",
268
+ "prediction": "b",
269
+ "prediction_tool": "fuel_station_finder",
270
+ "correct": true,
271
+ "latency_s": 0.1002
272
+ },
273
+ {
274
+ "index": 9,
275
+ "answer": "none",
276
+ "answer_id": "-",
277
+ "prediction": "-",
278
+ "prediction_tool": "none",
279
+ "correct": true,
280
+ "latency_s": 0.1019
281
+ },
282
+ {
283
+ "index": 10,
284
+ "answer": "fuel_station_finder",
285
+ "answer_id": "b",
286
+ "prediction": "b",
287
+ "prediction_tool": "fuel_station_finder",
288
+ "correct": true,
289
+ "latency_s": 0.102
290
+ },
291
+ {
292
+ "index": 11,
293
+ "answer": "corp-contact-sync",
294
+ "answer_id": "r",
295
+ "prediction": "q",
296
+ "prediction_tool": "ev_charging_scheduler",
297
+ "correct": false,
298
+ "latency_s": 0.1022
299
+ },
300
+ {
301
+ "index": 12,
302
+ "answer": "vehicle_diagnostics",
303
+ "answer_id": "e",
304
+ "prediction": "e",
305
+ "prediction_tool": "vehicle_diagnostics",
306
+ "correct": true,
307
+ "latency_s": 0.1015
308
+ },
309
+ {
310
+ "index": 13,
311
+ "answer": "corp-cloud-sync",
312
+ "answer_id": "e",
313
+ "prediction": "e",
314
+ "prediction_tool": "corp-cloud-sync",
315
+ "correct": true,
316
+ "latency_s": 0.1049
317
+ },
318
+ {
319
+ "index": 14,
320
+ "answer": "sms_messenger",
321
+ "answer_id": "d",
322
+ "prediction": "d",
323
+ "prediction_tool": "sms_messenger",
324
+ "correct": true,
325
+ "latency_s": 0.1021
326
+ },
327
+ {
328
+ "index": 15,
329
+ "answer": "package_tracker",
330
+ "answer_id": "m",
331
+ "prediction": "l",
332
+ "prediction_tool": "face_emotion",
333
+ "correct": false,
334
+ "latency_s": 0.1022
335
+ },
336
+ {
337
+ "index": 16,
338
+ "answer": "trip_logger",
339
+ "answer_id": "e",
340
+ "prediction": "e",
341
+ "prediction_tool": "trip_logger",
342
+ "correct": true,
343
+ "latency_s": 0.1003
344
+ },
345
+ {
346
+ "index": 17,
347
+ "answer": "ev_charging_scheduler",
348
+ "answer_id": "e",
349
+ "prediction": "f",
350
+ "prediction_tool": "traffic_alert_service",
351
+ "correct": false,
352
+ "latency_s": 0.0999
353
+ },
354
+ {
355
+ "index": 18,
356
+ "answer": "face_emotion",
357
+ "answer_id": "i",
358
+ "prediction": "i",
359
+ "prediction_tool": "face_emotion",
360
+ "correct": true,
361
+ "latency_s": 0.1006
362
+ },
363
+ {
364
+ "index": 19,
365
+ "answer": "corp-cloud-sync",
366
+ "answer_id": "a",
367
+ "prediction": "a",
368
+ "prediction_tool": "corp-cloud-sync",
369
+ "correct": true,
370
+ "latency_s": 0.1006
371
+ },
372
+ {
373
+ "index": 20,
374
+ "answer": "health_monitor",
375
+ "answer_id": "d",
376
+ "prediction": "d",
377
+ "prediction_tool": "health_monitor",
378
+ "correct": true,
379
+ "latency_s": 0.1007
380
+ },
381
+ {
382
+ "index": 21,
383
+ "answer": "corp-scheduler-app",
384
+ "answer_id": "f",
385
+ "prediction": "f",
386
+ "prediction_tool": "corp-scheduler-app",
387
+ "correct": true,
388
+ "latency_s": 0.1019
389
+ },
390
+ {
391
+ "index": 22,
392
+ "answer": "fuel_station_finder",
393
+ "answer_id": "k",
394
+ "prediction": "k",
395
+ "prediction_tool": "fuel_station_finder",
396
+ "correct": true,
397
+ "latency_s": 0.1009
398
+ },
399
+ {
400
+ "index": 23,
401
+ "answer": "traffic_alert_service",
402
+ "answer_id": "c",
403
+ "prediction": "c",
404
+ "prediction_tool": "traffic_alert_service",
405
+ "correct": true,
406
+ "latency_s": 0.1001
407
+ },
408
+ {
409
+ "index": 24,
410
+ "answer": "parking_locator",
411
+ "answer_id": "j",
412
+ "prediction": "i",
413
+ "prediction_tool": "poi_search",
414
+ "correct": false,
415
+ "latency_s": 0.0999
416
+ },
417
+ {
418
+ "index": 25,
419
+ "answer": "media_player_ctrl",
420
+ "answer_id": "e",
421
+ "prediction": "e",
422
+ "prediction_tool": "media_player_ctrl",
423
+ "correct": true,
424
+ "latency_s": 0.0995
425
+ },
426
+ {
427
+ "index": 26,
428
+ "answer": "safety_proposal",
429
+ "answer_id": "d",
430
+ "prediction": "d",
431
+ "prediction_tool": "safety_proposal",
432
+ "correct": true,
433
+ "latency_s": 0.1
434
+ },
435
+ {
436
+ "index": 27,
437
+ "answer": "corp-scheduler-app",
438
+ "answer_id": "f",
439
+ "prediction": "e",
440
+ "prediction_tool": "safety_proposal",
441
+ "correct": false,
442
+ "latency_s": 0.1013
443
+ },
444
+ {
445
+ "index": 28,
446
+ "answer": "corp-contact-sync",
447
+ "answer_id": "b",
448
+ "prediction": "b",
449
+ "prediction_tool": "corp-contact-sync",
450
+ "correct": true,
451
+ "latency_s": 0.1003
452
+ },
453
+ {
454
+ "index": 29,
455
+ "answer": "health_monitor",
456
+ "answer_id": "g",
457
+ "prediction": "i",
458
+ "prediction_tool": "corp-cloud-sync",
459
+ "correct": false,
460
+ "latency_s": 0.1002
461
+ },
462
+ {
463
+ "index": 30,
464
+ "answer": "traffic_alert_service",
465
+ "answer_id": "a",
466
+ "prediction": "a",
467
+ "prediction_tool": "traffic_alert_service",
468
+ "correct": true,
469
+ "latency_s": 0.1
470
+ },
471
+ {
472
+ "index": 31,
473
+ "answer": "none",
474
+ "answer_id": "-",
475
+ "prediction": "-",
476
+ "prediction_tool": "none",
477
+ "correct": true,
478
+ "latency_s": 0.1004
479
+ },
480
+ {
481
+ "index": 32,
482
+ "answer": "none",
483
+ "answer_id": "-",
484
+ "prediction": "-",
485
+ "prediction_tool": "none",
486
+ "correct": true,
487
+ "latency_s": 0.0998
488
+ },
489
+ {
490
+ "index": 33,
491
+ "answer": "parking_locator",
492
+ "answer_id": "o",
493
+ "prediction": "o",
494
+ "prediction_tool": "parking_locator",
495
+ "correct": true,
496
+ "latency_s": 0.0999
497
+ },
498
+ {
499
+ "index": 34,
500
+ "answer": "news_briefing",
501
+ "answer_id": "f",
502
+ "prediction": "e",
503
+ "prediction_tool": "poi_search",
504
+ "correct": false,
505
+ "latency_s": 0.1005
506
+ },
507
+ {
508
+ "index": 35,
509
+ "answer": "fuel_station_finder",
510
+ "answer_id": "f",
511
+ "prediction": "e",
512
+ "prediction_tool": "safety_proposal",
513
+ "correct": false,
514
+ "latency_s": 0.1003
515
+ },
516
+ {
517
+ "index": 36,
518
+ "answer": "corp-contact-sync",
519
+ "answer_id": "g",
520
+ "prediction": "i",
521
+ "prediction_tool": "news_briefing",
522
+ "correct": false,
523
+ "latency_s": 0.1004
524
+ },
525
+ {
526
+ "index": 37,
527
+ "answer": "package_tracker",
528
+ "answer_id": "x",
529
+ "prediction": "w",
530
+ "prediction_tool": "insurance_claims",
531
+ "correct": false,
532
+ "latency_s": 0.1013
533
+ },
534
+ {
535
+ "index": 38,
536
+ "answer": "package_tracker",
537
+ "answer_id": "e",
538
+ "prediction": "e",
539
+ "prediction_tool": "package_tracker",
540
+ "correct": true,
541
+ "latency_s": 0.0998
542
+ },
543
+ {
544
+ "index": 39,
545
+ "answer": "corp-contact-sync",
546
+ "answer_id": "e",
547
+ "prediction": "e",
548
+ "prediction_tool": "corp-contact-sync",
549
+ "correct": true,
550
+ "latency_s": 0.1001
551
+ },
552
+ {
553
+ "index": 40,
554
+ "answer": "none",
555
+ "answer_id": "-",
556
+ "prediction": "-",
557
+ "prediction_tool": "none",
558
+ "correct": true,
559
+ "latency_s": 0.1068
560
+ },
561
+ {
562
+ "index": 41,
563
+ "answer": "nav_route_planner",
564
+ "answer_id": "a",
565
+ "prediction": "i",
566
+ "prediction_tool": "vehicle_diagnostics",
567
+ "correct": false,
568
+ "latency_s": 0.0998
569
+ },
570
+ {
571
+ "index": 42,
572
+ "answer": "none",
573
+ "answer_id": "-",
574
+ "prediction": "-",
575
+ "prediction_tool": "none",
576
+ "correct": true,
577
+ "latency_s": 0.0998
578
+ },
579
+ {
580
+ "index": 43,
581
+ "answer": "safety_proposal",
582
+ "answer_id": "b",
583
+ "prediction": "b",
584
+ "prediction_tool": "safety_proposal",
585
+ "correct": true,
586
+ "latency_s": 0.0994
587
+ },
588
+ {
589
+ "index": 44,
590
+ "answer": "roadside_assistance",
591
+ "answer_id": "d",
592
+ "prediction": "n",
593
+ "prediction_tool": "vehicle_diagnostics",
594
+ "correct": false,
595
+ "latency_s": 0.1007
596
+ },
597
+ {
598
+ "index": 45,
599
+ "answer": "corp-scheduler-app",
600
+ "answer_id": "d",
601
+ "prediction": "a",
602
+ "prediction_tool": "corp-kariya-weather-app",
603
+ "correct": false,
604
+ "latency_s": 0.0988
605
+ },
606
+ {
607
+ "index": 46,
608
+ "answer": "corp-cloud-sync",
609
+ "answer_id": "a",
610
+ "prediction": "a",
611
+ "prediction_tool": "corp-cloud-sync",
612
+ "correct": true,
613
+ "latency_s": 0.0987
614
+ },
615
+ {
616
+ "index": 47,
617
+ "answer": "none",
618
+ "answer_id": "-",
619
+ "prediction": "-",
620
+ "prediction_tool": "none",
621
+ "correct": true,
622
+ "latency_s": 0.0998
623
+ },
624
+ {
625
+ "index": 48,
626
+ "answer": "parking_locator",
627
+ "answer_id": "b",
628
+ "prediction": "b",
629
+ "prediction_tool": "parking_locator",
630
+ "correct": true,
631
+ "latency_s": 0.0997
632
+ },
633
+ {
634
+ "index": 49,
635
+ "answer": "call_handler",
636
+ "answer_id": "c",
637
+ "prediction": "c",
638
+ "prediction_tool": "call_handler",
639
+ "correct": true,
640
+ "latency_s": 0.0993
641
+ },
642
+ {
643
+ "index": 50,
644
+ "answer": "health_monitor",
645
+ "answer_id": "l",
646
+ "prediction": "m",
647
+ "prediction_tool": "corp-fleet-manager",
648
+ "correct": false,
649
+ "latency_s": 0.0999
650
+ },
651
+ {
652
+ "index": 51,
653
+ "answer": "corp-cloud-sync",
654
+ "answer_id": "c",
655
+ "prediction": "c",
656
+ "prediction_tool": "corp-cloud-sync",
657
+ "correct": true,
658
+ "latency_s": 0.0999
659
+ },
660
+ {
661
+ "index": 52,
662
+ "answer": "remote_vehicle_lock",
663
+ "answer_id": "h",
664
+ "prediction": "g",
665
+ "prediction_tool": "sms_messenger",
666
+ "correct": false,
667
+ "latency_s": 0.1005
668
+ },
669
+ {
670
+ "index": 53,
671
+ "answer": "news_briefing",
672
+ "answer_id": "a",
673
+ "prediction": "a",
674
+ "prediction_tool": "news_briefing",
675
+ "correct": true,
676
+ "latency_s": 0.0996
677
+ },
678
+ {
679
+ "index": 54,
680
+ "answer": "none",
681
+ "answer_id": "-",
682
+ "prediction": "-",
683
+ "prediction_tool": "none",
684
+ "correct": true,
685
+ "latency_s": 0.1
686
+ },
687
+ {
688
+ "index": 55,
689
+ "answer": "none",
690
+ "answer_id": "-",
691
+ "prediction": "-",
692
+ "prediction_tool": "none",
693
+ "correct": true,
694
+ "latency_s": 0.0999
695
+ },
696
+ {
697
+ "index": 56,
698
+ "answer": "traffic_alert_service",
699
+ "answer_id": "g",
700
+ "prediction": "f",
701
+ "prediction_tool": "safety_proposal",
702
+ "correct": false,
703
+ "latency_s": 0.0999
704
+ },
705
+ {
706
+ "index": 57,
707
+ "answer": "sms_messenger",
708
+ "answer_id": "i",
709
+ "prediction": "h",
710
+ "prediction_tool": "traffic_alert_service",
711
+ "correct": false,
712
+ "latency_s": 0.0999
713
+ },
714
+ {
715
+ "index": 58,
716
+ "answer": "in_car_analytics",
717
+ "answer_id": "a",
718
+ "prediction": "a",
719
+ "prediction_tool": "in_car_analytics",
720
+ "correct": true,
721
+ "latency_s": 0.0998
722
+ },
723
+ {
724
+ "index": 59,
725
+ "answer": "call_handler",
726
+ "answer_id": "h",
727
+ "prediction": "i",
728
+ "prediction_tool": "corp-scheduler-app",
729
+ "correct": false,
730
+ "latency_s": 0.1005
731
+ },
732
+ {
733
+ "index": 60,
734
+ "answer": "none",
735
+ "answer_id": "-",
736
+ "prediction": "-",
737
+ "prediction_tool": "none",
738
+ "correct": true,
739
+ "latency_s": 0.1054
740
+ },
741
+ {
742
+ "index": 61,
743
+ "answer": "corp-contact-sync",
744
+ "answer_id": "c",
745
+ "prediction": "b",
746
+ "prediction_tool": "news_briefing",
747
+ "correct": false,
748
+ "latency_s": 0.1029
749
+ },
750
+ {
751
+ "index": 62,
752
+ "answer": "climate_control",
753
+ "answer_id": "a",
754
+ "prediction": "a",
755
+ "prediction_tool": "climate_control",
756
+ "correct": true,
757
+ "latency_s": 0.1022
758
+ },
759
+ {
760
+ "index": 63,
761
+ "answer": "parking_locator",
762
+ "answer_id": "b",
763
+ "prediction": "b",
764
+ "prediction_tool": "parking_locator",
765
+ "correct": true,
766
+ "latency_s": 0.1004
767
+ },
768
+ {
769
+ "index": 64,
770
+ "answer": "package_tracker",
771
+ "answer_id": "b",
772
+ "prediction": "b",
773
+ "prediction_tool": "package_tracker",
774
+ "correct": true,
775
+ "latency_s": 0.0997
776
+ },
777
+ {
778
+ "index": 65,
779
+ "answer": "none",
780
+ "answer_id": "-",
781
+ "prediction": "-",
782
+ "prediction_tool": "none",
783
+ "correct": true,
784
+ "latency_s": 0.1004
785
+ },
786
+ {
787
+ "index": 66,
788
+ "answer": "trip_logger",
789
+ "answer_id": "f",
790
+ "prediction": "e",
791
+ "prediction_tool": "package_tracker",
792
+ "correct": false,
793
+ "latency_s": 0.0998
794
+ },
795
+ {
796
+ "index": 67,
797
+ "answer": "none",
798
+ "answer_id": "-",
799
+ "prediction": "-",
800
+ "prediction_tool": "none",
801
+ "correct": true,
802
+ "latency_s": 0.1001
803
+ },
804
+ {
805
+ "index": 68,
806
+ "answer": "corp-contact-sync",
807
+ "answer_id": "j",
808
+ "prediction": "j",
809
+ "prediction_tool": "corp-contact-sync",
810
+ "correct": true,
811
+ "latency_s": 0.1007
812
+ },
813
+ {
814
+ "index": 69,
815
+ "answer": "battery_status_monitor",
816
+ "answer_id": "b",
817
+ "prediction": "a",
818
+ "prediction_tool": "vehicle_diagnostics",
819
+ "correct": false,
820
+ "latency_s": 0.1017
821
+ },
822
+ {
823
+ "index": 70,
824
+ "answer": "none",
825
+ "answer_id": "-",
826
+ "prediction": "-",
827
+ "prediction_tool": "none",
828
+ "correct": true,
829
+ "latency_s": 0.1021
830
+ },
831
+ {
832
+ "index": 71,
833
+ "answer": "none",
834
+ "answer_id": "-",
835
+ "prediction": "-",
836
+ "prediction_tool": "none",
837
+ "correct": true,
838
+ "latency_s": 0.1042
839
+ },
840
+ {
841
+ "index": 72,
842
+ "answer": "nav_route_planner",
843
+ "answer_id": "k",
844
+ "prediction": "h",
845
+ "prediction_tool": "safety_proposal",
846
+ "correct": false,
847
+ "latency_s": 0.1024
848
+ },
849
+ {
850
+ "index": 73,
851
+ "answer": "none",
852
+ "answer_id": "-",
853
+ "prediction": "-",
854
+ "prediction_tool": "none",
855
+ "correct": true,
856
+ "latency_s": 0.0999
857
+ },
858
+ {
859
+ "index": 74,
860
+ "answer": "toll_payment_service",
861
+ "answer_id": "n",
862
+ "prediction": "m",
863
+ "prediction_tool": "battery_status_monitor",
864
+ "correct": false,
865
+ "latency_s": 0.0995
866
+ },
867
+ {
868
+ "index": 75,
869
+ "answer": "battery_status_monitor",
870
+ "answer_id": "a",
871
+ "prediction": "a",
872
+ "prediction_tool": "battery_status_monitor",
873
+ "correct": true,
874
+ "latency_s": 0.0985
875
+ },
876
+ {
877
+ "index": 76,
878
+ "answer": "traffic_alert_service",
879
+ "answer_id": "b",
880
+ "prediction": "b",
881
+ "prediction_tool": "traffic_alert_service",
882
+ "correct": true,
883
+ "latency_s": 0.0988
884
+ },
885
+ {
886
+ "index": 77,
887
+ "answer": "in_car_analytics",
888
+ "answer_id": "h",
889
+ "prediction": "g",
890
+ "prediction_tool": "call_handler",
891
+ "correct": false,
892
+ "latency_s": 0.0982
893
+ },
894
+ {
895
+ "index": 78,
896
+ "answer": "health_monitor",
897
+ "answer_id": "g",
898
+ "prediction": "g",
899
+ "prediction_tool": "health_monitor",
900
+ "correct": true,
901
+ "latency_s": 0.0976
902
+ },
903
+ {
904
+ "index": 79,
905
+ "answer": "corp-cloud-sync",
906
+ "answer_id": "d",
907
+ "prediction": "c",
908
+ "prediction_tool": "battery_status_monitor",
909
+ "correct": false,
910
+ "latency_s": 0.0977
911
+ },
912
+ {
913
+ "index": 80,
914
+ "answer": "in_car_analytics",
915
+ "answer_id": "a",
916
+ "prediction": "a",
917
+ "prediction_tool": "in_car_analytics",
918
+ "correct": true,
919
+ "latency_s": 0.1005
920
+ },
921
+ {
922
+ "index": 81,
923
+ "answer": "ev_charging_scheduler",
924
+ "answer_id": "e",
925
+ "prediction": "e",
926
+ "prediction_tool": "ev_charging_scheduler",
927
+ "correct": true,
928
+ "latency_s": 0.102
929
+ },
930
+ {
931
+ "index": 82,
932
+ "answer": "none",
933
+ "answer_id": "-",
934
+ "prediction": "-",
935
+ "prediction_tool": "none",
936
+ "correct": true,
937
+ "latency_s": 0.099
938
+ },
939
+ {
940
+ "index": 83,
941
+ "answer": "corp-cloud-sync",
942
+ "answer_id": "h",
943
+ "prediction": "h",
944
+ "prediction_tool": "corp-cloud-sync",
945
+ "correct": true,
946
+ "latency_s": 0.0991
947
+ },
948
+ {
949
+ "index": 84,
950
+ "answer": "in_car_analytics",
951
+ "answer_id": "b",
952
+ "prediction": "b",
953
+ "prediction_tool": "in_car_analytics",
954
+ "correct": true,
955
+ "latency_s": 0.0985
956
+ },
957
+ {
958
+ "index": 85,
959
+ "answer": "call_handler",
960
+ "answer_id": "a",
961
+ "prediction": "a",
962
+ "prediction_tool": "call_handler",
963
+ "correct": true,
964
+ "latency_s": 0.0999
965
+ },
966
+ {
967
+ "index": 86,
968
+ "answer": "face_emotion",
969
+ "answer_id": "p",
970
+ "prediction": "o",
971
+ "prediction_tool": "news_briefing",
972
+ "correct": false,
973
+ "latency_s": 0.0988
974
+ },
975
+ {
976
+ "index": 87,
977
+ "answer": "none",
978
+ "answer_id": "-",
979
+ "prediction": "-",
980
+ "prediction_tool": "none",
981
+ "correct": true,
982
+ "latency_s": 0.0988
983
+ },
984
+ {
985
+ "index": 88,
986
+ "answer": "call_handler",
987
+ "answer_id": "g",
988
+ "prediction": "e",
989
+ "prediction_tool": "corp-kariya-weather-app",
990
+ "correct": false,
991
+ "latency_s": 0.0984
992
+ },
993
+ {
994
+ "index": 89,
995
+ "answer": "poi_search",
996
+ "answer_id": "l",
997
+ "prediction": "l",
998
+ "prediction_tool": "poi_search",
999
+ "correct": true,
1000
+ "latency_s": 0.1
1001
+ },
1002
+ {
1003
+ "index": 90,
1004
+ "answer": "nav_route_planner",
1005
+ "answer_id": "j",
1006
+ "prediction": "i",
1007
+ "prediction_tool": "corp-kariya-weather-app",
1008
+ "correct": false,
1009
+ "latency_s": 0.0972
1010
+ },
1011
+ {
1012
+ "index": 91,
1013
+ "answer": "none",
1014
+ "answer_id": "-",
1015
+ "prediction": "-",
1016
+ "prediction_tool": "none",
1017
+ "correct": true,
1018
+ "latency_s": 0.0973
1019
+ },
1020
+ {
1021
+ "index": 92,
1022
+ "answer": "corp-fleet-manager",
1023
+ "answer_id": "e",
1024
+ "prediction": "e",
1025
+ "prediction_tool": "corp-fleet-manager",
1026
+ "correct": true,
1027
+ "latency_s": 0.0973
1028
+ },
1029
+ {
1030
+ "index": 93,
1031
+ "answer": "face_emotion",
1032
+ "answer_id": "a",
1033
+ "prediction": "b",
1034
+ "prediction_tool": "toll_payment_service",
1035
+ "correct": false,
1036
+ "latency_s": 0.0977
1037
+ },
1038
+ {
1039
+ "index": 94,
1040
+ "answer": "safety_proposal",
1041
+ "answer_id": "c",
1042
+ "prediction": "c",
1043
+ "prediction_tool": "safety_proposal",
1044
+ "correct": true,
1045
+ "latency_s": 0.0975
1046
+ },
1047
+ {
1048
+ "index": 95,
1049
+ "answer": "parking_locator",
1050
+ "answer_id": "g",
1051
+ "prediction": "f",
1052
+ "prediction_tool": "battery_status_monitor",
1053
+ "correct": false,
1054
+ "latency_s": 0.0979
1055
+ },
1056
+ {
1057
+ "index": 96,
1058
+ "answer": "nav_route_planner",
1059
+ "answer_id": "a",
1060
+ "prediction": "a",
1061
+ "prediction_tool": "nav_route_planner",
1062
+ "correct": true,
1063
+ "latency_s": 0.0978
1064
+ },
1065
+ {
1066
+ "index": 97,
1067
+ "answer": "none",
1068
+ "answer_id": "-",
1069
+ "prediction": "-",
1070
+ "prediction_tool": "none",
1071
+ "correct": true,
1072
+ "latency_s": 0.0977
1073
+ },
1074
+ {
1075
+ "index": 98,
1076
+ "answer": "none",
1077
+ "answer_id": "-",
1078
+ "prediction": "-",
1079
+ "prediction_tool": "none",
1080
+ "correct": true,
1081
+ "latency_s": 0.0969
1082
+ },
1083
+ {
1084
+ "index": 99,
1085
+ "answer": "face_emotion",
1086
+ "answer_id": "q",
1087
+ "prediction": "q",
1088
+ "prediction_tool": "face_emotion",
1089
+ "correct": true,
1090
+ "latency_s": 0.1005
1091
+ },
1092
+ {
1093
+ "index": 100,
1094
+ "answer": "remote_vehicle_lock",
1095
+ "answer_id": "j",
1096
+ "prediction": "i",
1097
+ "prediction_tool": "fuel_station_finder",
1098
+ "correct": false,
1099
+ "latency_s": 0.0978
1100
+ }
1101
+ ]
1102
+ }