Repository navigation
Expand file tree
/
Copy pathtoolmap.json
More file actions
3028 lines (3028 loc) · 101 KB
/
Copy pathtoolmap.json
File metadata and controls
3028 lines (3028 loc) · 101 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
{
"schema": "jev-match-bench toolmap v2",
"generated": "2026-10-12",
"generated_from": [
"results/route1/score.md",
"results/ex1/score.md",
"results/pv1/score-de-adjudicated.md",
"results/pv1/score-en.md",
"reference/pv1/adjudicated-a.de.jsonl",
"reference/pv1/claude-c.en.jsonl",
"result rows for versions, dates and Jev tokens",
"plan/toolmap-plan.yaml",
"results/route1/automation.json",
"results/pv1/automation.json",
"results/capabilities.json"
],
"findings": {
"route1": "results/route1/findings.md",
"ex1": "results/ex1/findings.md",
"pv1": "results/pv1/gliner.md",
"automation": "results/automation-findings.md",
"c1": "results/c1/findings.md",
"c2": "results/c2/findings.md",
"c3": "results/c3/findings.md",
"c4": "results/c4/findings.md",
"c5": "results/c5/findings.md",
"c6": "results/c6/findings.md",
"c7": "results/c7/findings.md",
"c8": "results/c8/findings.md",
"c9": "results/c9/findings.md"
},
"note": "Each cell is one task and one tool family. \"best_arm\" is the arm of that family with the highest value on the same test data the value is measured on, so it is chosen after the fact and flatters the families with many arms (GLiNER has up to six arms in route1 and nine in pv1, the local family has one arm in route1 and ex1 and four in pv1). Cells are per task: a family that is best on one task says nothing about another. All numbers are run 1, one run per arm. For German rows, English-only GLiNER models (GLiNER2.5-Decide, GLiNER2.5-Decide-1B) are not eligible as best arm, in route1 and pv1 alike; their best value is listed as english_model_extra. pv1 rows: baseline computed from the reference with the rule of results/pv1/gliner.md, run dates 2026-10-01 (Jev, local) and 2026-10-08 (GLiNER). skill = (value - baseline) / (100 - baseline). p50 values in route1 rows are per question group (the p50 ms column of results/route1/score.md); the Speed table of results/route1/findings.md gives medians per source and language, so the two can differ by a few ms. Latencies are not one clock: Jev is an HTTP round trip to api.typesafe.ai, the local LLM an HTTP call over the LAN, GLiNER GPU time in-process. Local cost (hardware, power) was not estimated. Tests: route1 and pv1 cells carry top_family (the family with the highest value in the row) and mcnemar_vs_top_p, an exact two-sided McNemar between this family's best arm and the top family's best arm on the same items (unanswered counts as wrong); route1 and pv1 cells other than Jev also carry mcnemar_vs_jev_p. Both compared arms are picked after the fact as the best of their family, and no p value is corrected for that or for the number of tests, so p >= 0.05 means the difference is not established, not that the two are level. ex1 cells have no test; their skill is F1 against a baseline F1 of 0, exact_frame_skill compares whole frames against extracting nothing.",
"classic_note": "The classic family (plan/t3.md) is a trained classifier: multilingual-e5-base sentence embeddings plus logistic regression, fitted on the MASSIVE 1.1 train split of the same locale. The other three families get no training data. top_family and mcnemar_vs_top_p are therefore computed among jev, gliner and local-llm only; a classic cell carries top_family (that zero-shot top), its own mcnemar_vs_top_p against it, and mcnemar_vs_jev_p. Its two arms are classic-e5-lr (full train split) and classic-e5-lr-10shot (10 rows per class); best_arm is chosen as for every family.",
"local_series_note": "T12 (plan/t12.md): route1 and ex1 rows carry local_size_classes, one column per size class, each the one Qwen3 model named for that class before any T12 run (small qwen3-4b, medium qwen3-14b, large qwen3-32b, moe qwen3-30b-a3b). They are not best-of picks. The local-llm family cell stays restricted to winnow-12b in route1 and ex1, so no existing cell changed. Every other local model of the series is reported only in results/t12-findings.md. local_size_contrast is the one preregistered test per route1 row (large against small).",
"states": {
"measured": "a value from a run; see the cell's source_report",
"planned": "a dated test design is committed (see plan); nothing is measured yet",
"open": "a sensible test that nobody has designed yet",
"not_applicable": "the tool cannot do the task; see reason"
},
"families": {
"jev": "TypeSafe Jev, hosted API (api.typesafe.ai)",
"gliner": "Fastino GLiNER2 / GLiNER2.5 encoders, local GPU (RTX 4090)",
"local-llm": "open LLM through Ollama, local GPU (RTX 4090)",
"classic": "sentence embeddings plus a trained logistic regression, local (Mac, MPS)"
},
"tools": {
"jev": {
"name": "Jev (jev-latest, measured as jev-1.13.0)",
"maker": "TypeSafe AI, Inc., San Francisco, CA",
"runs": "hosted API (api.typesafe.ai). TypeSafe's privacy policy says the services are hosted in the United States.",
"licence": "proprietary hosted service, no weights or code published; use is governed by TypeSafe's terms",
"primitives": "choice, score, noul",
"sources": [
"https://typesafe.ai/legal/privacy-policy (section International Visitors, last updated 2025-11-19, read 2026-10-09)",
"https://typesafe.ai/legal/terms (offered by TypeSafe AI, Inc., 255 California St, San Francisco; last updated 2026-09-19, read 2026-10-09)",
"https://docs.typesafe.ai/introduction (three primitives choice, score, noul, read 2026-10-09)"
]
},
"gliner": {
"name": "GLiNER2 / GLiNER2.5 (Decide classifiers and entity extractors)",
"maker": "Fastino",
"runs": "local, Python package gliner2 with the weights from Hugging Face, on one RTX 4090",
"licence": "Apache-2.0, for the gliner2 library and for every model measured here (GLiNER2.5-multi-Decide, GLiNER2.5-Decide, GLiNER2.5-Decide-1B, gliner2.5-multi-v1, gliner2-large-v1)",
"sources": [
"https://github.com/fastino-ai/GLiNER2 (licence Apache-2.0, read 2026-10-09)",
"https://huggingface.co/fastino (model cards, license apache-2.0 in each card's metadata, read 2026-10-09)"
]
},
"local-llm": {
"name": "open LLM through Ollama; the arm measured in route1 and ex1 is Winnow-12B (Q8_0)",
"maker": "EldanRing (Winnow-12B, a Gemma 4 12B IT with a merged LoRA). pv1 also measured Qwen3 32B, Qwen3.8 27B and shisa-de-1, see results/pv1/method.md",
"runs": "local, Ollama on one RTX 4090",
"licence": "Apache-2.0 for Winnow-12B (license field of its model card; base model google/gemma-4-12B-it). Other local models carry their own licences",
"sources": [
"https://huggingface.co/EldanRing/Winnow-12B (license apache-2.0, base_model google/gemma-4-12B-it, read 2026-10-09)"
]
},
"classic": {
"name": "sentence embeddings (intfloat/multilingual-e5-base) plus a scikit-learn LogisticRegression trained on labelled examples",
"maker": "Microsoft (multilingual-e5-base, published as intfloat on Hugging Face, Wang et al., arXiv:2402.05672) and the scikit-learn developers (LogisticRegression); the training here is ours, see plan/t3.md",
"runs": "local, Python with sentence-transformers and scikit-learn, on a Mac (Apple silicon, MPS); no GPU server needed",
"licence": "MIT for multilingual-e5-base (card metadata), BSD-3-Clause for scikit-learn, Apache-2.0 for sentence-transformers. The trained classifier inherits the licence of its training data (MASSIVE, CC BY 4.0)",
"sources": [
"https://huggingface.co/intfloat/multilingual-e5-base (license mit, revision d128750597153bb5987e10b1c3493a34e5a4502a, read 2026-10-09)",
"https://github.com/scikit-learn/scikit-learn (licence BSD-3-Clause, read 2026-10-09)",
"https://github.com/UKPLab/sentence-transformers (licence Apache-2.0, read 2026-10-09)"
]
}
},
"rows": [
{
"id": "route1-scenario-de",
"kind": "route",
"task": "route: MASSIVE scenario (18 classes)",
"lang": "de",
"metric": "accuracy",
"source_report": "results/route1/score.md",
"section": "massive scenario, de",
"cells": {
"jev": {
"state": "measured",
"best_arm": "jev",
"tied_with": [],
"arms_compared": 1,
"value": 72.8,
"macro_f1": 73.3,
"baseline": 13.5,
"baseline_name": "most frequent answer",
"skill": 0.686,
"n": 2974,
"answered": 2974,
"mcnemar_vs_jev_p": null,
"p50_ms_per_decision": 232,
"measured": "2026-10-09",
"model_version": "jev-1.13.0",
"source_report": "results/route1/score.md",
"median_input_tokens_per_decision": 510,
"usd_per_1000_decisions": 0.0214,
"cost_note": "median input tokens per call x USD 0.042 per million input tokens, output free (docs.typesafe.ai/models, read 2026-10-02); per source and language, not per question group",
"top_family": "jev",
"mcnemar_vs_top_p": null
},
"gliner": {
"state": "measured",
"best_arm": "gliner2.5-multi-decide",
"tied_with": [],
"arms_compared": 2,
"value": 48.5,
"macro_f1": 46.4,
"baseline": 13.5,
"baseline_name": "most frequent answer",
"skill": 0.404,
"n": 2974,
"answered": 2974,
"mcnemar_vs_jev_p": "<0.001",
"p50_ms_per_decision": 14,
"measured": "2026-10-09",
"model_version": "fastino/GLiNER2.5-multi-Decide (revision e173bca1f0c4)",
"source_report": "results/route1/score.md",
"english_model_extra": {
"arm": "gliner2.5-decide",
"tied_with": [],
"value": 49,
"skill": 0.41,
"note": "English-only model on German rows, labelled extra in the report, not eligible as best arm"
},
"top_family": "jev",
"mcnemar_vs_top_p": "<0.001",
"vs_top_paired": 2974,
"vs_top_only_top_right": 912,
"vs_top_only_arm_right": 188
},
"local-llm": {
"state": "measured",
"best_arm": "winnow-12b",
"tied_with": [],
"arms_compared": 1,
"value": 71.8,
"macro_f1": 72.3,
"baseline": 13.5,
"baseline_name": "most frequent answer",
"skill": 0.674,
"n": 2974,
"answered": 2974,
"mcnemar_vs_jev_p": 0.11,
"p50_ms_per_decision": 89,
"measured": "2026-10-09",
"model_version": "hf.co/EldanRing/Winnow-12B:Q8_0 (Ollama digest b5bedb246945, ollama 0.35.0)",
"source_report": "results/route1/score.md",
"top_family": "jev",
"mcnemar_vs_top_p": 0.11,
"vs_top_paired": 2974,
"vs_top_only_top_right": 192,
"vs_top_only_arm_right": 161
},
"classic": {
"state": "measured",
"best_arm": "classic-e5-lr",
"tied_with": [],
"arms_compared": 2,
"value": 86.9,
"macro_f1": 86.1,
"baseline": 13.5,
"baseline_name": "most frequent answer",
"skill": 0.848,
"n": 2974,
"answered": 2974,
"mcnemar_vs_jev_p": "<0.001",
"p50_ms_per_decision": 25,
"measured": "2026-10-09",
"model_version": "intfloat/multilingual-e5-base + sklearn LogisticRegression(C=1.0, multinomial) (revision d12875059715)",
"source_report": "results/route1/score.md",
"top_family": "jev",
"mcnemar_vs_top_p": "<0.001",
"vs_top_paired": 2974,
"vs_top_only_top_right": 141,
"vs_top_only_arm_right": 559
}
},
"local_size_classes": {
"small": {
"arm": "qwen3-4b",
"value": 57.5,
"macro_f1": 59.3,
"baseline": 13.5,
"baseline_name": "most frequent answer",
"skill": 0.509,
"n": 2974,
"answered": 2974,
"mcnemar_vs_jev_p": "<0.001",
"p50_ms_per_decision": 36,
"letter_mass_median": 1,
"measured": "2026-10-09",
"model_version": "qwen3:4b-q4_K_M (Ollama digest 2bfd38a7daaf, ollama 0.35.0)",
"source_report": "results/route1/score.md"
},
"medium": {
"arm": "qwen3-14b",
"value": 65.6,
"macro_f1": 66.4,
"baseline": 13.5,
"baseline_name": "most frequent answer",
"skill": 0.602,
"n": 2974,
"answered": 2974,
"mcnemar_vs_jev_p": "<0.001",
"p50_ms_per_decision": 52,
"letter_mass_median": 1,
"measured": "2026-10-09",
"model_version": "qwen3:14b (Ollama digest bdbd181c33f2, ollama 0.35.0)",
"source_report": "results/route1/score.md"
},
"large": {
"arm": "qwen3-32b",
"value": 67.3,
"macro_f1": 69.7,
"baseline": 13.5,
"baseline_name": "most frequent answer",
"skill": 0.622,
"n": 2974,
"answered": 2974,
"mcnemar_vs_jev_p": "<0.001",
"p50_ms_per_decision": 160,
"letter_mass_median": 1,
"measured": "2026-10-09",
"model_version": "qwen3:32b (Ollama digest 030ee887880f, ollama 0.35.0)",
"source_report": "results/route1/score.md"
},
"moe": {
"arm": "qwen3-30b-a3b",
"value": 59.8,
"macro_f1": 61.6,
"baseline": 13.5,
"baseline_name": "most frequent answer",
"skill": 0.535,
"n": 2974,
"answered": 2974,
"mcnemar_vs_jev_p": "<0.001",
"p50_ms_per_decision": 66,
"letter_mass_median": 0.8789,
"measured": "2026-10-09",
"model_version": "qwen3:30b-a3b-q4_K_M (Ollama digest 0b28110b7a33, ollama 0.35.0)",
"source_report": "results/route1/score.md"
}
},
"local_size_contrast": {
"test": "exact two-sided McNemar, large against small representative, preregistered in plan/t12.md",
"large": "qwen3-32b",
"small": "qwen3-4b",
"paired": 2974,
"only_large_right": 459,
"only_small_right": 168,
"p": "<0.001"
}
},
{
"id": "route1-scenario-en",
"kind": "route",
"task": "route: MASSIVE scenario (18 classes)",
"lang": "en",
"metric": "accuracy",
"source_report": "results/route1/score.md",
"section": "massive scenario, en",
"cells": {
"jev": {
"state": "measured",
"best_arm": "jev",
"tied_with": [],
"arms_compared": 1,
"value": 74.1,
"macro_f1": 74.2,
"baseline": 13.5,
"baseline_name": "most frequent answer",
"skill": 0.701,
"n": 2974,
"answered": 2974,
"mcnemar_vs_jev_p": null,
"p50_ms_per_decision": 231,
"measured": "2026-10-09",
"model_version": "jev-1.13.0",
"source_report": "results/route1/score.md",
"median_input_tokens_per_decision": 504,
"usd_per_1000_decisions": 0.0212,
"cost_note": "median input tokens per call x USD 0.042 per million input tokens, output free (docs.typesafe.ai/models, read 2026-10-02); per source and language, not per question group",
"top_family": "jev",
"mcnemar_vs_top_p": null
},
"gliner": {
"state": "measured",
"best_arm": "gliner2.5-decide",
"tied_with": [],
"arms_compared": 6,
"value": 60.9,
"macro_f1": 60.2,
"baseline": 13.5,
"baseline_name": "most frequent answer",
"skill": 0.547,
"n": 2974,
"answered": 2974,
"mcnemar_vs_jev_p": "<0.001",
"p50_ms_per_decision": 26,
"measured": "2026-10-09",
"model_version": "fastino/GLiNER2.5-Decide (revision 9d1bfb848cd1)",
"source_report": "results/route1/score.md",
"top_family": "jev",
"mcnemar_vs_top_p": "<0.001",
"vs_top_paired": 2974,
"vs_top_only_top_right": 543,
"vs_top_only_arm_right": 149
},
"local-llm": {
"state": "measured",
"best_arm": "winnow-12b",
"tied_with": [],
"arms_compared": 1,
"value": 70.6,
"macro_f1": 71.8,
"baseline": 13.5,
"baseline_name": "most frequent answer",
"skill": 0.661,
"n": 2974,
"answered": 2974,
"mcnemar_vs_jev_p": "<0.001",
"p50_ms_per_decision": 92,
"measured": "2026-10-09",
"model_version": "hf.co/EldanRing/Winnow-12B:Q8_0 (Ollama digest b5bedb246945, ollama 0.35.0)",
"source_report": "results/route1/score.md",
"top_family": "jev",
"mcnemar_vs_top_p": "<0.001",
"vs_top_paired": 2974,
"vs_top_only_top_right": 237,
"vs_top_only_arm_right": 134
},
"classic": {
"state": "measured",
"best_arm": "classic-e5-lr",
"tied_with": [],
"arms_compared": 2,
"value": 89.8,
"macro_f1": 89.6,
"baseline": 13.5,
"baseline_name": "most frequent answer",
"skill": 0.883,
"n": 2974,
"answered": 2974,
"mcnemar_vs_jev_p": "<0.001",
"p50_ms_per_decision": 23,
"measured": "2026-10-09",
"model_version": "intfloat/multilingual-e5-base + sklearn LogisticRegression(C=1.0, multinomial) (revision d12875059715)",
"source_report": "results/route1/score.md",
"top_family": "jev",
"mcnemar_vs_top_p": "<0.001",
"vs_top_paired": 2974,
"vs_top_only_top_right": 92,
"vs_top_only_arm_right": 560
}
},
"local_size_classes": {
"small": {
"arm": "qwen3-4b",
"value": 63,
"macro_f1": 65.2,
"baseline": 13.5,
"baseline_name": "most frequent answer",
"skill": 0.572,
"n": 2974,
"answered": 2974,
"mcnemar_vs_jev_p": "<0.001",
"p50_ms_per_decision": 36,
"letter_mass_median": 1,
"measured": "2026-10-09",
"model_version": "qwen3:4b-q4_K_M (Ollama digest 2bfd38a7daaf, ollama 0.35.0)",
"source_report": "results/route1/score.md"
},
"medium": {
"arm": "qwen3-14b",
"value": 69,
"macro_f1": 71,
"baseline": 13.5,
"baseline_name": "most frequent answer",
"skill": 0.642,
"n": 2974,
"answered": 2974,
"mcnemar_vs_jev_p": "<0.001",
"p50_ms_per_decision": 50,
"letter_mass_median": 1,
"measured": "2026-10-09",
"model_version": "qwen3:14b (Ollama digest bdbd181c33f2, ollama 0.35.0)",
"source_report": "results/route1/score.md"
},
"large": {
"arm": "qwen3-32b",
"value": 69.8,
"macro_f1": 72.9,
"baseline": 13.5,
"baseline_name": "most frequent answer",
"skill": 0.651,
"n": 2974,
"answered": 2974,
"mcnemar_vs_jev_p": "<0.001",
"p50_ms_per_decision": 163,
"letter_mass_median": 1,
"measured": "2026-10-09",
"model_version": "qwen3:32b (Ollama digest 030ee887880f, ollama 0.35.0)",
"source_report": "results/route1/score.md"
},
"moe": {
"arm": "qwen3-30b-a3b",
"value": 61.5,
"macro_f1": 62.6,
"baseline": 13.5,
"baseline_name": "most frequent answer",
"skill": 0.554,
"n": 2974,
"answered": 2974,
"mcnemar_vs_jev_p": "<0.001",
"p50_ms_per_decision": 61,
"letter_mass_median": 0.8789,
"measured": "2026-10-09",
"model_version": "qwen3:30b-a3b-q4_K_M (Ollama digest 0b28110b7a33, ollama 0.35.0)",
"source_report": "results/route1/score.md"
}
},
"local_size_contrast": {
"test": "exact two-sided McNemar, large against small representative, preregistered in plan/t12.md",
"large": "qwen3-32b",
"small": "qwen3-4b",
"paired": 2974,
"only_large_right": 368,
"only_small_right": 166,
"p": "<0.001"
}
},
{
"id": "route1-intent2-de",
"kind": "route",
"task": "route: MASSIVE intent given the gold scenario, 2+ options",
"lang": "de",
"metric": "accuracy",
"source_report": "results/route1/score.md",
"section": "massive intent, 2+ options, de",
"cells": {
"jev": {
"state": "measured",
"best_arm": "jev",
"tied_with": [],
"arms_compared": 1,
"value": 90.7,
"macro_f1": 86.5,
"baseline": 50.2,
"baseline_name": "most frequent answer",
"skill": 0.812,
"n": 2622,
"answered": 2622,
"mcnemar_vs_jev_p": null,
"p50_ms_per_decision": 234,
"measured": "2026-10-09",
"model_version": "jev-1.13.0",
"source_report": "results/route1/score.md",
"median_input_tokens_per_decision": 510,
"usd_per_1000_decisions": 0.0214,
"cost_note": "median input tokens per call x USD 0.042 per million input tokens, output free (docs.typesafe.ai/models, read 2026-10-02); per source and language, not per question group",
"top_family": "jev",
"mcnemar_vs_top_p": null
},
"gliner": {
"state": "measured",
"best_arm": "gliner2.5-multi-decide",
"tied_with": [],
"arms_compared": 2,
"value": 58.9,
"macro_f1": 54.9,
"baseline": 50.2,
"baseline_name": "most frequent answer",
"skill": 0.175,
"n": 2622,
"answered": 2622,
"mcnemar_vs_jev_p": "<0.001",
"p50_ms_per_decision": 13,
"measured": "2026-10-09",
"model_version": "fastino/GLiNER2.5-multi-Decide (revision e173bca1f0c4)",
"source_report": "results/route1/score.md",
"english_model_extra": {
"arm": "gliner2.5-decide",
"tied_with": [],
"value": 65.9,
"skill": 0.316,
"note": "English-only model on German rows, labelled extra in the report, not eligible as best arm"
},
"top_family": "jev",
"mcnemar_vs_top_p": "<0.001",
"vs_top_paired": 2622,
"vs_top_only_top_right": 899,
"vs_top_only_arm_right": 67
},
"local-llm": {
"state": "measured",
"best_arm": "winnow-12b",
"tied_with": [],
"arms_compared": 1,
"value": 89,
"macro_f1": 82.5,
"baseline": 50.2,
"baseline_name": "most frequent answer",
"skill": 0.779,
"n": 2622,
"answered": 2621,
"mcnemar_vs_jev_p": 0.002,
"p50_ms_per_decision": 84,
"measured": "2026-10-09",
"model_version": "hf.co/EldanRing/Winnow-12B:Q8_0 (Ollama digest b5bedb246945, ollama 0.35.0)",
"source_report": "results/route1/score.md",
"top_family": "jev",
"mcnemar_vs_top_p": 0.002,
"vs_top_paired": 2622,
"vs_top_only_top_right": 115,
"vs_top_only_arm_right": 71
},
"classic": {
"state": "measured",
"best_arm": "classic-e5-lr",
"tied_with": [],
"arms_compared": 2,
"value": 86.9,
"macro_f1": 72.5,
"baseline": 50.2,
"baseline_name": "most frequent answer",
"skill": 0.737,
"n": 2622,
"answered": 2622,
"mcnemar_vs_jev_p": "<0.001",
"p50_ms_per_decision": 24,
"measured": "2026-10-09",
"model_version": "intfloat/multilingual-e5-base + sklearn LogisticRegression(C=1.0, multinomial) (revision d12875059715)",
"source_report": "results/route1/score.md",
"top_family": "jev",
"mcnemar_vs_top_p": "<0.001",
"vs_top_paired": 2622,
"vs_top_only_top_right": 243,
"vs_top_only_arm_right": 144
}
},
"local_size_classes": {
"small": {
"arm": "qwen3-4b",
"value": 77,
"macro_f1": 72.1,
"baseline": 50.2,
"baseline_name": "most frequent answer",
"skill": 0.539,
"n": 2622,
"answered": 2622,
"mcnemar_vs_jev_p": "<0.001",
"p50_ms_per_decision": 35,
"letter_mass_median": 1,
"measured": "2026-10-09",
"model_version": "qwen3:4b-q4_K_M (Ollama digest 2bfd38a7daaf, ollama 0.35.0)",
"source_report": "results/route1/score.md"
},
"medium": {
"arm": "qwen3-14b",
"value": 79.9,
"macro_f1": 78.1,
"baseline": 50.2,
"baseline_name": "most frequent answer",
"skill": 0.596,
"n": 2622,
"answered": 2622,
"mcnemar_vs_jev_p": "<0.001",
"p50_ms_per_decision": 46,
"letter_mass_median": 1,
"measured": "2026-10-09",
"model_version": "qwen3:14b (Ollama digest bdbd181c33f2, ollama 0.35.0)",
"source_report": "results/route1/score.md"
},
"large": {
"arm": "qwen3-32b",
"value": 84.2,
"macro_f1": 83.1,
"baseline": 50.2,
"baseline_name": "most frequent answer",
"skill": 0.682,
"n": 2622,
"answered": 2622,
"mcnemar_vs_jev_p": "<0.001",
"p50_ms_per_decision": 144.5,
"letter_mass_median": 1,
"measured": "2026-10-09",
"model_version": "qwen3:32b (Ollama digest 030ee887880f, ollama 0.35.0)",
"source_report": "results/route1/score.md"
},
"moe": {
"arm": "qwen3-30b-a3b",
"value": 70.8,
"macro_f1": 67.1,
"baseline": 50.2,
"baseline_name": "most frequent answer",
"skill": 0.413,
"n": 2622,
"answered": 2622,
"mcnemar_vs_jev_p": "<0.001",
"p50_ms_per_decision": 52,
"letter_mass_median": 0.8789,
"measured": "2026-10-09",
"model_version": "qwen3:30b-a3b-q4_K_M (Ollama digest 0b28110b7a33, ollama 0.35.0)",
"source_report": "results/route1/score.md"
}
},
"local_size_contrast": {
"test": "exact two-sided McNemar, large against small representative, preregistered in plan/t12.md",
"large": "qwen3-32b",
"small": "qwen3-4b",
"paired": 2622,
"only_large_right": 316,
"only_small_right": 129,
"p": "<0.001"
}
},
{
"id": "route1-intent2-en",
"kind": "route",
"task": "route: MASSIVE intent given the gold scenario, 2+ options",
"lang": "en",
"metric": "accuracy",
"source_report": "results/route1/score.md",
"section": "massive intent, 2+ options, en",
"cells": {
"jev": {
"state": "measured",
"best_arm": "jev",
"tied_with": [],
"arms_compared": 1,
"value": 92,
"macro_f1": 88.4,
"baseline": 50.2,
"baseline_name": "most frequent answer",
"skill": 0.838,
"n": 2622,
"answered": 2622,
"mcnemar_vs_jev_p": null,
"p50_ms_per_decision": 230,
"measured": "2026-10-09",
"model_version": "jev-1.13.0",
"source_report": "results/route1/score.md",
"median_input_tokens_per_decision": 504,
"usd_per_1000_decisions": 0.0212,
"cost_note": "median input tokens per call x USD 0.042 per million input tokens, output free (docs.typesafe.ai/models, read 2026-10-02); per source and language, not per question group",
"top_family": "jev",
"mcnemar_vs_top_p": null
},
"gliner": {
"state": "measured",
"best_arm": "gliner2.5-decide",
"tied_with": [],
"arms_compared": 6,
"value": 78.3,
"macro_f1": 74.3,
"baseline": 50.2,
"baseline_name": "most frequent answer",
"skill": 0.565,
"n": 2622,
"answered": 2622,
"mcnemar_vs_jev_p": "<0.001",
"p50_ms_per_decision": 26,
"measured": "2026-10-09",
"model_version": "fastino/GLiNER2.5-Decide (revision 9d1bfb848cd1)",
"source_report": "results/route1/score.md",
"top_family": "jev",
"mcnemar_vs_top_p": "<0.001",
"vs_top_paired": 2622,
"vs_top_only_top_right": 428,
"vs_top_only_arm_right": 71
},
"local-llm": {
"state": "measured",
"best_arm": "winnow-12b",
"tied_with": [],
"arms_compared": 1,
"value": 91.6,
"macro_f1": 86.4,
"baseline": 50.2,
"baseline_name": "most frequent answer",
"skill": 0.832,
"n": 2622,
"answered": 2622,
"mcnemar_vs_jev_p": 0.575,
"p50_ms_per_decision": 88,
"measured": "2026-10-09",
"model_version": "hf.co/EldanRing/Winnow-12B:Q8_0 (Ollama digest b5bedb246945, ollama 0.35.0)",
"source_report": "results/route1/score.md",
"top_family": "jev",
"mcnemar_vs_top_p": 0.575,
"vs_top_paired": 2622,
"vs_top_only_top_right": 82,
"vs_top_only_arm_right": 74
},
"classic": {
"state": "measured",
"best_arm": "classic-e5-lr",
"tied_with": [],
"arms_compared": 2,
"value": 91.2,
"macro_f1": 79.6,
"baseline": 50.2,
"baseline_name": "most frequent answer",
"skill": 0.824,
"n": 2622,
"answered": 2622,
"mcnemar_vs_jev_p": 0.298,
"p50_ms_per_decision": 23,
"measured": "2026-10-09",
"model_version": "intfloat/multilingual-e5-base + sklearn LogisticRegression(C=1.0, multinomial) (revision d12875059715)",
"source_report": "results/route1/score.md",
"top_family": "jev",
"mcnemar_vs_top_p": 0.298,
"vs_top_paired": 2622,
"vs_top_only_top_right": 159,
"vs_top_only_arm_right": 140
}
},
"local_size_classes": {
"small": {
"arm": "qwen3-4b",
"value": 82.6,
"macro_f1": 76.1,
"baseline": 50.2,
"baseline_name": "most frequent answer",
"skill": 0.65,
"n": 2622,
"answered": 2622,
"mcnemar_vs_jev_p": "<0.001",
"p50_ms_per_decision": 35,
"letter_mass_median": 1,
"measured": "2026-10-09",
"model_version": "qwen3:4b-q4_K_M (Ollama digest 2bfd38a7daaf, ollama 0.35.0)",
"source_report": "results/route1/score.md"
},
"medium": {
"arm": "qwen3-14b",
"value": 83.5,
"macro_f1": 81.3,
"baseline": 50.2,
"baseline_name": "most frequent answer",
"skill": 0.669,
"n": 2622,
"answered": 2622,
"mcnemar_vs_jev_p": "<0.001",
"p50_ms_per_decision": 45,
"letter_mass_median": 1,
"measured": "2026-10-09",
"model_version": "qwen3:14b (Ollama digest bdbd181c33f2, ollama 0.35.0)",
"source_report": "results/route1/score.md"
},
"large": {
"arm": "qwen3-32b",
"value": 86.7,
"macro_f1": 85.4,
"baseline": 50.2,
"baseline_name": "most frequent answer",
"skill": 0.732,
"n": 2622,
"answered": 2622,
"mcnemar_vs_jev_p": "<0.001",
"p50_ms_per_decision": 148,
"letter_mass_median": 1,
"measured": "2026-10-09",
"model_version": "qwen3:32b (Ollama digest 030ee887880f, ollama 0.35.0)",
"source_report": "results/route1/score.md"
},
"moe": {
"arm": "qwen3-30b-a3b",
"value": 77,
"macro_f1": 74.1,
"baseline": 50.2,
"baseline_name": "most frequent answer",
"skill": 0.539,
"n": 2622,
"answered": 2622,
"mcnemar_vs_jev_p": "<0.001",
"p50_ms_per_decision": 49,
"letter_mass_median": 0.8789,
"measured": "2026-10-09",
"model_version": "qwen3:30b-a3b-q4_K_M (Ollama digest 0b28110b7a33, ollama 0.35.0)",
"source_report": "results/route1/score.md"
}
},
"local_size_contrast": {
"test": "exact two-sided McNemar, large against small representative, preregistered in plan/t12.md",
"large": "qwen3-32b",
"small": "qwen3-4b",
"paired": 2622,
"only_large_right": 229,
"only_small_right": 122,
"p": "<0.001"
}
},
{
"id": "route1-fastdec26-en",
"kind": "route",
"task": "route: fast-decisions dev, single-label tasks with 26 labels or fewer",
"lang": "en",
"metric": "accuracy",
"source_report": "results/route1/score.md",
"section": "fastdec, tasks with 26 options or fewer, en",
"cells": {
"jev": {
"state": "measured",
"best_arm": "jev",
"tied_with": [],
"arms_compared": 1,
"value": 63,
"macro_f1": 55.9,
"baseline": 31.2,
"baseline_name": "most frequent answer",
"skill": 0.461,
"n": 2500,
"answered": 2500,
"mcnemar_vs_jev_p": null,
"p50_ms_per_decision": 240,
"measured": "2026-10-09",
"model_version": "jev-1.13.0",
"source_report": "results/route1/score.md",
"median_input_tokens_per_decision": 544,
"usd_per_1000_decisions": 0.0228,
"cost_note": "median input tokens per call x USD 0.042 per million input tokens, output free (docs.typesafe.ai/models, read 2026-10-02); per source and language, not per question group",
"top_family": "gliner",
"mcnemar_vs_top_p": 0.168,
"vs_top_paired": 2500,
"vs_top_only_top_right": 341,
"vs_top_only_arm_right": 305
},
"gliner": {
"state": "measured",
"best_arm": "gliner2.5-decide-1b-intext",
"tied_with": [],
"arms_compared": 6,
"value": 64.4,
"macro_f1": 60.4,
"baseline": 31.2,
"baseline_name": "most frequent answer",
"skill": 0.482,
"n": 2500,
"answered": 2500,
"mcnemar_vs_jev_p": 0.168,
"p50_ms_per_decision": 26,
"measured": "2026-10-09",
"model_version": "fastino/GLiNER2.5-Decide-1B (revision 67af949e17be)",
"source_report": "results/route1/score.md",
"top_family": "gliner",
"mcnemar_vs_top_p": null
},
"local-llm": {
"state": "measured",
"best_arm": "winnow-12b",
"tied_with": [],
"arms_compared": 1,
"value": 63.6,
"macro_f1": 55.3,
"baseline": 31.2,
"baseline_name": "most frequent answer",
"skill": 0.471,
"n": 2500,
"answered": 2500,
"mcnemar_vs_jev_p": 0.309,
"p50_ms_per_decision": 120,
"measured": "2026-10-09",
"model_version": "hf.co/EldanRing/Winnow-12B:Q8_0 (Ollama digest b5bedb246945, ollama 0.35.0)",
"source_report": "results/route1/score.md",
"top_family": "gliner",
"mcnemar_vs_top_p": 0.483,
"vs_top_paired": 2500,
"vs_top_only_top_right": 338,
"vs_top_only_arm_right": 319
},
"classic": {
"state": "not_applicable",
"reason": "fast-decisions publishes only the dev split, no train split, so the classic classifier has nothing to train on (plan/t3.md)",
"plan": "plan/t3.md"
}
},
"note": "Home ground for GLiNER: Fastino built fast-decisions, and its dataset card calls it the classification suite behind GLiNER2.5-Decide. Whether these dev rows were used in training is not stated, so a Decide lead or tie here is not a neutral result.",
"local_size_classes": {
"small": {
"arm": "qwen3-4b",
"value": 59.7,
"macro_f1": 54.2,
"baseline": 31.2,
"baseline_name": "most frequent answer",
"skill": 0.414,
"n": 2500,
"answered": 2500,
"mcnemar_vs_jev_p": "<0.001",
"p50_ms_per_decision": 43,
"letter_mass_median": 1,
"measured": "2026-10-09",
"model_version": "qwen3:4b-q4_K_M (Ollama digest 2bfd38a7daaf, ollama 0.35.0)",
"source_report": "results/route1/score.md"
},
"medium": {
"arm": "qwen3-14b",
"value": 60.9,
"macro_f1": 54.9,
"baseline": 31.2,
"baseline_name": "most frequent answer",
"skill": 0.432,
"n": 2500,
"answered": 2500,
"mcnemar_vs_jev_p": 0.005,
"p50_ms_per_decision": 64,
"letter_mass_median": 1,
"measured": "2026-10-09",
"model_version": "qwen3:14b (Ollama digest bdbd181c33f2, ollama 0.35.0)",
"source_report": "results/route1/score.md"
},
"large": {
"arm": "qwen3-32b",
"value": 60.6,
"macro_f1": 54,
"baseline": 31.2,
"baseline_name": "most frequent answer",
"skill": 0.428,
"n": 2500,
"answered": 2500,
"mcnemar_vs_jev_p": 0.002,
"p50_ms_per_decision": 200,
"letter_mass_median": 1,
"measured": "2026-10-09",
"model_version": "qwen3:32b (Ollama digest 030ee887880f, ollama 0.35.0)",
"source_report": "results/route1/score.md"
},
"moe": {
"arm": "qwen3-30b-a3b",
"value": 56.2,
"macro_f1": 47.5,
"baseline": 31.2,
"baseline_name": "most frequent answer",
"skill": 0.364,
"n": 2500,
"answered": 2500,
"mcnemar_vs_jev_p": "<0.001",
"p50_ms_per_decision": 65,
"letter_mass_median": 0.8789,
"measured": "2026-10-09",
"model_version": "qwen3:30b-a3b-q4_K_M (Ollama digest 0b28110b7a33, ollama 0.35.0)",
"source_report": "results/route1/score.md"
}
},
"local_size_contrast": {
"test": "exact two-sided McNemar, large against small representative, preregistered in plan/t12.md",
"large": "qwen3-32b",
"small": "qwen3-4b",
"paired": 2500,
"only_large_right": 192,
"only_small_right": 168,
"p": 0.225
}
},
{
"id": "ex1-slots-de",
"kind": "extract",
"task": "slot extraction (MASSIVE slots)",
"lang": "de",
"metric": "micro F1, threshold 0.5 for GLiNER",
"source_report": "results/ex1/score.md",
"section": "de",
"cells": {
"jev": {
"state": "not_applicable",
"reason": "Jev has no extraction primitive (choice, score, noul only), docs.typesafe.ai/introduction",
"source_report": "results/ex1/method.md"
},
"gliner": {
"state": "measured",
"best_arm": "gliner2.5-multi",
"tied_with": [],
"arms_compared": 1,
"value": 22.5,
"precision": 28.8,
"recall": 18.5,