guychuk commited on
Commit
970f2df
·
verified ·
1 Parent(s): 3559cba

typical-medium (tm1b): weights, evals, inference package, manifest

Browse files
MANIFEST.json ADDED
@@ -0,0 +1,12 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "release": "typical-medium",
3
+ "checkpoint": "tm1b",
4
+ "backbone": "Qwen/Qwen3-4B-Base",
5
+ "tap_layer": 26,
6
+ "sha256_best_pt": "0500c3faeff1b24a3c7e56a846346d9e33d88c8906c48c976f1ad19e95a7e14e",
7
+ "source_commit": "2fad326c4bcffec4dbc4e8341f9c8248a13f65ae",
8
+ "date": "2026-09-21",
9
+ "parent": "typical-small",
10
+ "private_source": "guychuk/pcdm-runs/typical-medium",
11
+ "public_repo": "OzLabs/typical-medium"
12
+ }
bench.json ADDED
@@ -0,0 +1,425 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "model": "runs/tm1b",
3
+ "energy_model": null,
4
+ "backbone": "Qwen/Qwen3-4B-Base",
5
+ "quick": false,
6
+ "b_chunk": 4,
7
+ "native": {
8
+ "256": {
9
+ "2": {
10
+ "state_tokens": 256,
11
+ "suffix_tokens": 22,
12
+ "nc_chunk": 8,
13
+ "native": [
14
+ {
15
+ "m": 1,
16
+ "t_state_ms": 17.69845560193062,
17
+ "t_queries_ms": 38.907185196876526,
18
+ "total_ms": 56.605640798807144,
19
+ "ratio": 1.0,
20
+ "marginal_ms": 38.907185196876526
21
+ },
22
+ {
23
+ "m": 32,
24
+ "t_state_ms": 17.29227602481842,
25
+ "t_queries_ms": 107.14780166745186,
26
+ "total_ms": 124.44007769227028,
27
+ "ratio": 2.1983688539904773,
28
+ "marginal_ms": 3.3483688021078706
29
+ },
30
+ {
31
+ "m": 256,
32
+ "t_state_ms": 17.395589500665665,
33
+ "t_queries_ms": 720.7270376384258,
34
+ "total_ms": 738.1226271390915,
35
+ "ratio": 13.039736265200023,
36
+ "marginal_ms": 2.815339990775101
37
+ }
38
+ ],
39
+ "native_peak_mem_mb": 14565.473792,
40
+ "b_batched": [
41
+ {
42
+ "m": 1,
43
+ "t_state_ms": 19.00634914636612,
44
+ "t_queries_ms": 65.67803397774696,
45
+ "total_ms": 84.68438312411308,
46
+ "ratio": 1.0,
47
+ "marginal_ms": 65.67803397774696
48
+ },
49
+ {
50
+ "m": 32,
51
+ "t_state_ms": 19.073322415351868,
52
+ "t_queries_ms": 408.54376927018166,
53
+ "total_ms": 427.6170916855335,
54
+ "ratio": 5.0495389576000065,
55
+ "marginal_ms": 12.766992789693177
56
+ },
57
+ {
58
+ "m": 256,
59
+ "t_state_ms": 19.45146918296814,
60
+ "t_queries_ms": 3124.6551759541035,
61
+ "total_ms": 3144.1066451370716,
62
+ "ratio": 37.127348976836515,
63
+ "marginal_ms": 12.205684281070717
64
+ }
65
+ ],
66
+ "b_batched_peak_mem_mb": 15050.985472
67
+ },
68
+ "4": {
69
+ "state_tokens": 256,
70
+ "suffix_tokens": 34,
71
+ "nc_chunk": 8,
72
+ "native": [
73
+ {
74
+ "m": 1,
75
+ "t_state_ms": 17.520517110824585,
76
+ "t_queries_ms": 39.11682963371277,
77
+ "total_ms": 56.63734674453735,
78
+ "ratio": 1.0,
79
+ "marginal_ms": 39.11682963371277
80
+ },
81
+ {
82
+ "m": 32,
83
+ "t_state_ms": 17.56681129336357,
84
+ "t_queries_ms": 109.09038409590721,
85
+ "total_ms": 126.65719538927078,
86
+ "ratio": 2.2362840540634403,
87
+ "marginal_ms": 3.4090745029971004
88
+ },
89
+ {
90
+ "m": 256,
91
+ "t_state_ms": 17.606686800718307,
92
+ "t_queries_ms": 820.20478323102,
93
+ "total_ms": 837.8114700317383,
94
+ "ratio": 14.792562120021005,
95
+ "marginal_ms": 3.2039249344961718
96
+ }
97
+ ],
98
+ "native_peak_mem_mb": 14579.467776,
99
+ "b_batched": [
100
+ {
101
+ "m": 1,
102
+ "t_state_ms": 19.45377141237259,
103
+ "t_queries_ms": 66.85380637645721,
104
+ "total_ms": 86.3075777888298,
105
+ "ratio": 1.0,
106
+ "marginal_ms": 66.85380637645721
107
+ },
108
+ {
109
+ "m": 32,
110
+ "t_state_ms": 19.49465647339821,
111
+ "t_queries_ms": 415.58846831321716,
112
+ "total_ms": 435.0831247866154,
113
+ "ratio": 5.041076762125575,
114
+ "marginal_ms": 12.987139634788036
115
+ },
116
+ {
117
+ "m": 256,
118
+ "t_state_ms": 19.42378655076027,
119
+ "t_queries_ms": 3201.4211788773537,
120
+ "total_ms": 3220.844965428114,
121
+ "ratio": 37.31821756495831,
122
+ "marginal_ms": 12.505551479989663
123
+ }
124
+ ],
125
+ "b_batched_peak_mem_mb": 15418.156544
126
+ },
127
+ "10": {
128
+ "state_tokens": 256,
129
+ "suffix_tokens": 70,
130
+ "nc_chunk": 8,
131
+ "native": [
132
+ {
133
+ "m": 1,
134
+ "t_state_ms": 17.71542802453041,
135
+ "t_queries_ms": 39.49485346674919,
136
+ "total_ms": 57.2102814912796,
137
+ "ratio": 1.0,
138
+ "marginal_ms": 39.49485346674919
139
+ },
140
+ {
141
+ "m": 32,
142
+ "t_state_ms": 17.667438834905624,
143
+ "t_queries_ms": 113.24474588036537,
144
+ "total_ms": 130.912184715271,
145
+ "ratio": 2.288263250989694,
146
+ "marginal_ms": 3.538898308761418
147
+ },
148
+ {
149
+ "m": 256,
150
+ "t_state_ms": 17.80020445585251,
151
+ "t_queries_ms": 772.7814465761185,
152
+ "total_ms": 790.581651031971,
153
+ "ratio": 13.818873643411054,
154
+ "marginal_ms": 3.0186775256879628
155
+ }
156
+ ],
157
+ "native_peak_mem_mb": 14627.14112,
158
+ "b_batched": [
159
+ {
160
+ "m": 1,
161
+ "t_state_ms": 19.491098821163177,
162
+ "t_queries_ms": 67.33075901865959,
163
+ "total_ms": 86.82185783982277,
164
+ "ratio": 1.0,
165
+ "marginal_ms": 67.33075901865959
166
+ },
167
+ {
168
+ "m": 32,
169
+ "t_state_ms": 19.476331770420074,
170
+ "t_queries_ms": 518.6272598803043,
171
+ "total_ms": 538.1035916507244,
172
+ "ratio": 6.19778941661752,
173
+ "marginal_ms": 16.20710187125951
174
+ },
175
+ {
176
+ "m": 256,
177
+ "t_state_ms": 19.5954367518425,
178
+ "t_queries_ms": 4071.4379996061325,
179
+ "total_ms": 4091.033436357975,
180
+ "ratio": 47.11985596882185,
181
+ "marginal_ms": 15.904054685961455
182
+ }
183
+ ],
184
+ "b_batched_peak_mem_mb": 16549.823488
185
+ },
186
+ "32": {
187
+ "state_tokens": 256,
188
+ "suffix_tokens": 224,
189
+ "nc_chunk": 8,
190
+ "native": [
191
+ {
192
+ "m": 1,
193
+ "t_state_ms": 17.933987081050873,
194
+ "t_queries_ms": 40.07749259471893,
195
+ "total_ms": 58.011479675769806,
196
+ "ratio": 1.0,
197
+ "marginal_ms": 40.07749259471893
198
+ },
199
+ {
200
+ "m": 32,
201
+ "t_state_ms": 17.875414341688156,
202
+ "t_queries_ms": 199.15443286299706,
203
+ "total_ms": 217.0298472046852,
204
+ "ratio": 3.74115344786377,
205
+ "marginal_ms": 6.223576026968658
206
+ },
207
+ {
208
+ "m": 256,
209
+ "t_state_ms": 17.83185824751854,
210
+ "t_queries_ms": 1439.1961731016636,
211
+ "total_ms": 1457.0280313491821,
212
+ "ratio": 25.116201818891934,
213
+ "marginal_ms": 5.621860051178373
214
+ }
215
+ ],
216
+ "native_peak_mem_mb": 14892.906496,
217
+ "b_batched": [
218
+ {
219
+ "m": 1,
220
+ "t_state_ms": 19.53354850411415,
221
+ "t_queries_ms": 75.62482729554176,
222
+ "total_ms": 95.15837579965591,
223
+ "ratio": 1.0,
224
+ "marginal_ms": 75.62482729554176
225
+ },
226
+ {
227
+ "m": 32,
228
+ "t_state_ms": 19.454356282949448,
229
+ "t_queries_ms": 903.2446146011353,
230
+ "total_ms": 922.6989708840847,
231
+ "ratio": 9.696455652277129,
232
+ "marginal_ms": 28.226394206285477
233
+ },
234
+ {
235
+ "m": 256,
236
+ "t_state_ms": 21.368220448493958,
237
+ "t_queries_ms": 7146.388553082943,
238
+ "total_ms": 7167.756773531437,
239
+ "ratio": 75.32449680123014,
240
+ "marginal_ms": 27.915580285480246
241
+ }
242
+ ],
243
+ "b_batched_peak_mem_mb": 20708.114944
244
+ },
245
+ "64": {
246
+ "state_tokens": 256,
247
+ "suffix_tokens": 460,
248
+ "nc_chunk": 8,
249
+ "native": [
250
+ {
251
+ "m": 1,
252
+ "t_state_ms": 17.79353618621826,
253
+ "t_queries_ms": 41.496291756629944,
254
+ "total_ms": 59.289827942848206,
255
+ "ratio": 1.0,
256
+ "marginal_ms": 41.496291756629944
257
+ },
258
+ {
259
+ "m": 32,
260
+ "t_state_ms": 17.75621995329857,
261
+ "t_queries_ms": 338.5215550661087,
262
+ "total_ms": 356.2777750194073,
263
+ "ratio": 6.009087686387578,
264
+ "marginal_ms": 10.578798595815897
265
+ },
266
+ {
267
+ "m": 256,
268
+ "t_state_ms": 17.848066985607147,
269
+ "t_queries_ms": 2598.344925791025,
270
+ "total_ms": 2616.1929927766323,
271
+ "ratio": 44.1254947694988,
272
+ "marginal_ms": 10.149784866371192
273
+ }
274
+ ],
275
+ "native_peak_mem_mb": 15333.215232,
276
+ "b_batched": [
277
+ {
278
+ "m": 1,
279
+ "t_state_ms": 19.58613097667694,
280
+ "t_queries_ms": 93.24939921498299,
281
+ "total_ms": 112.83553019165993,
282
+ "ratio": 1.0,
283
+ "marginal_ms": 93.24939921498299
284
+ },
285
+ {
286
+ "m": 32,
287
+ "t_state_ms": 19.635286182165146,
288
+ "t_queries_ms": 1480.6604906916618,
289
+ "total_ms": 1500.295776873827,
290
+ "ratio": 13.296306352488953,
291
+ "marginal_ms": 46.27064033411443
292
+ },
293
+ {
294
+ "m": 256,
295
+ "t_state_ms": 19.47050541639328,
296
+ "t_queries_ms": 11805.44163286686,
297
+ "total_ms": 11824.912138283253,
298
+ "ratio": 104.79777174971146,
299
+ "marginal_ms": 46.11500637838617
300
+ }
301
+ ],
302
+ "b_batched_peak_mem_mb": 26712.366592
303
+ },
304
+ "128": {
305
+ "state_tokens": 256,
306
+ "suffix_tokens": 1029,
307
+ "nc_chunk": 8,
308
+ "native": [
309
+ {
310
+ "m": 1,
311
+ "t_state_ms": 17.73315668106079,
312
+ "t_queries_ms": 48.95908385515213,
313
+ "total_ms": 66.69224053621292,
314
+ "ratio": 1.0,
315
+ "marginal_ms": 48.95908385515213
316
+ },
317
+ {
318
+ "m": 32,
319
+ "t_state_ms": 17.543818801641464,
320
+ "t_queries_ms": 716.6986837983131,
321
+ "total_ms": 734.2425025999546,
322
+ "ratio": 11.00941423914633,
323
+ "marginal_ms": 22.396833868697286
324
+ },
325
+ {
326
+ "m": 256,
327
+ "t_state_ms": 17.649024724960327,
328
+ "t_queries_ms": 5805.374853312969,
329
+ "total_ms": 5823.0238780379295,
330
+ "ratio": 87.31186463702792,
331
+ "marginal_ms": 22.677245520753786
332
+ }
333
+ ],
334
+ "native_peak_mem_mb": 16306.219008,
335
+ "b_batched": [
336
+ {
337
+ "m": 1,
338
+ "t_state_ms": 25.08896216750145,
339
+ "t_queries_ms": 140.0025337934494,
340
+ "total_ms": 165.09149596095085,
341
+ "ratio": 1.0,
342
+ "marginal_ms": 140.0025337934494
343
+ },
344
+ {
345
+ "m": 32,
346
+ "t_state_ms": 19.624758511781693,
347
+ "t_queries_ms": 2846.7205613851547,
348
+ "total_ms": 2866.3453198969364,
349
+ "ratio": 17.362162134474293,
350
+ "marginal_ms": 88.96001754328609
351
+ },
352
+ {
353
+ "m": 256,
354
+ "t_state_ms": 19.967511296272278,
355
+ "t_queries_ms": 22693.809997290373,
356
+ "total_ms": 22713.777508586645,
357
+ "ratio": 137.58296498784614,
358
+ "marginal_ms": 88.64769530191552
359
+ }
360
+ ],
361
+ "b_batched_peak_mem_mb": 33338.615296
362
+ },
363
+ "256": {
364
+ "state_tokens": 256,
365
+ "suffix_tokens": 2309,
366
+ "nc_chunk": 8,
367
+ "native": [
368
+ {
369
+ "m": 1,
370
+ "t_state_ms": 17.903797328472137,
371
+ "t_queries_ms": 78.47601175308228,
372
+ "total_ms": 96.37980908155441,
373
+ "ratio": 1.0,
374
+ "marginal_ms": 78.47601175308228
375
+ },
376
+ {
377
+ "m": 32,
378
+ "t_state_ms": 17.928332090377808,
379
+ "t_queries_ms": 1699.1598010063171,
380
+ "total_ms": 1717.088133096695,
381
+ "ratio": 17.81584908145786,
382
+ "marginal_ms": 53.09874378144741
383
+ },
384
+ {
385
+ "m": 256,
386
+ "t_state_ms": 17.86353439092636,
387
+ "t_queries_ms": 12893.904343247414,
388
+ "total_ms": 12911.76787763834,
389
+ "ratio": 133.9675602253237,
390
+ "marginal_ms": 50.36681384081021
391
+ }
392
+ ],
393
+ "native_peak_mem_mb": 18598.0416,
394
+ "b_batched": [
395
+ {
396
+ "m": 1,
397
+ "t_state_ms": 20.501967519521713,
398
+ "t_queries_ms": 222.51605615019798,
399
+ "total_ms": 243.0180236697197,
400
+ "ratio": 1.0,
401
+ "marginal_ms": 222.51605615019798
402
+ },
403
+ {
404
+ "m": 32,
405
+ "t_state_ms": 19.544314593076706,
406
+ "t_queries_ms": 5423.821311444044,
407
+ "total_ms": 5443.365626037121,
408
+ "ratio": 22.399020220142503,
409
+ "marginal_ms": 169.49441598262638
410
+ },
411
+ {
412
+ "m": 256,
413
+ "t_state_ms": 19.62793990969658,
414
+ "t_queries_ms": 43251.52888149023,
415
+ "total_ms": 43271.15682139993,
416
+ "ratio": 178.0573974225417,
417
+ "marginal_ms": 168.9512846933212
418
+ }
419
+ ],
420
+ "b_batched_peak_mem_mb": 27778.753536
421
+ }
422
+ }
423
+ },
424
+ "crossover": {}
425
+ }
best.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0500c3faeff1b24a3c7e56a846346d9e33d88c8906c48c976f1ad19e95a7e14e
3
+ size 46546609
eval_wf.json ADDED
@@ -0,0 +1,3292 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "run": "runs/tm1b",
3
+ "mode": "native",
4
+ "max_state": 4096,
5
+ "limit": 0,
6
+ "eval": {
7
+ "u_chaosnli": {
8
+ "n": 1594,
9
+ "n_total": 1594,
10
+ "suffix_overflow": 0,
11
+ "raw": {
12
+ "nll": 1.014103315474928,
13
+ "brier": 0.17091631474795344,
14
+ "jsd": 0.06646093306121999,
15
+ "acc": 0.5131744040150564,
16
+ "acc_k": 0.5131744040150564,
17
+ "ece": 0.10623594570713328,
18
+ "sel_acc80": 0.5313479623824452,
19
+ "coverage": 1.0,
20
+ "false_abstain": 0.0,
21
+ "null_recall": "n/a"
22
+ },
23
+ "by_family": {
24
+ "": {
25
+ "nll": 1.014103315474928,
26
+ "brier": 0.17091631474795344,
27
+ "jsd": 0.06646093306121999,
28
+ "acc": 0.5131744040150564,
29
+ "acc_k": 0.5131744040150564,
30
+ "ece": 0.10623594570713328,
31
+ "sel_acc80": 0.5313479623824452,
32
+ "coverage": 1.0,
33
+ "false_abstain": 0.0,
34
+ "null_recall": "n/a",
35
+ "n": 1594
36
+ }
37
+ },
38
+ "by_qtype": {},
39
+ "by_label_class": {
40
+ "answer": {
41
+ "nll": 1.014103315474928,
42
+ "brier": 0.17091631474795344,
43
+ "jsd": 0.06646093306121999,
44
+ "acc": 0.5131744040150564,
45
+ "acc_k": 0.5131744040150564,
46
+ "ece": 0.10623594570713328,
47
+ "sel_acc80": 0.5313479623824452,
48
+ "coverage": 1.0,
49
+ "false_abstain": 0.0,
50
+ "null_recall": "n/a",
51
+ "n": 1594
52
+ }
53
+ },
54
+ "baselines": {
55
+ "majority_class": 0,
56
+ "majority_acc": 0.46361355081555833
57
+ }
58
+ },
59
+ "u_real_heldout": {
60
+ "n": 562,
61
+ "n_total": 562,
62
+ "suffix_overflow": 0,
63
+ "raw": {
64
+ "nll": 0.6395687143934158,
65
+ "brier": 0.11684180639496176,
66
+ "jsd": 0.05530236546295147,
67
+ "acc": 0.6583629893238434,
68
+ "acc_k": 0.6583629893238434,
69
+ "ece": 0.04802694863696116,
70
+ "sel_acc80": 0.7,
71
+ "coverage": 1.0,
72
+ "false_abstain": 0.0,
73
+ "null_recall": "n/a"
74
+ },
75
+ "by_family": {
76
+ "": {
77
+ "nll": 0.6395687143934158,
78
+ "brier": 0.11684180639496176,
79
+ "jsd": 0.05530236546295147,
80
+ "acc": 0.6583629893238434,
81
+ "acc_k": 0.6583629893238434,
82
+ "ece": 0.04802694863696116,
83
+ "sel_acc80": 0.7,
84
+ "coverage": 1.0,
85
+ "false_abstain": 0.0,
86
+ "null_recall": "n/a",
87
+ "n": 562
88
+ }
89
+ },
90
+ "by_qtype": {},
91
+ "by_label_class": {
92
+ "answer": {
93
+ "nll": 0.6395687143934158,
94
+ "brier": 0.11684180639496176,
95
+ "jsd": 0.05530236546295147,
96
+ "acc": 0.6583629893238434,
97
+ "acc_k": 0.6583629893238434,
98
+ "ece": 0.04802694863696116,
99
+ "sel_acc80": 0.7,
100
+ "coverage": 1.0,
101
+ "false_abstain": 0.0,
102
+ "null_recall": "n/a",
103
+ "n": 562
104
+ }
105
+ },
106
+ "baselines": {
107
+ "majority_class": 0,
108
+ "majority_acc": 0.5071174377224199
109
+ }
110
+ },
111
+ "u_synthetic_heldout": {
112
+ "n": 475,
113
+ "n_total": 475,
114
+ "suffix_overflow": 0,
115
+ "raw": {
116
+ "nll": 0.6619388394682302,
117
+ "brier": 0.029335292543417674,
118
+ "jsd": 0.01347952047234738,
119
+ "acc": 0.9073684210526316,
120
+ "acc_k": 0.9073684210526316,
121
+ "ece": 0.17639653883482279,
122
+ "sel_acc80": 0.9815789473684211,
123
+ "coverage": 1.0,
124
+ "false_abstain": 0.0,
125
+ "null_recall": "n/a"
126
+ },
127
+ "by_family": {
128
+ "": {
129
+ "nll": 0.6619388394682302,
130
+ "brier": 0.029335292543417674,
131
+ "jsd": 0.01347952047234738,
132
+ "acc": 0.9073684210526316,
133
+ "acc_k": 0.9073684210526316,
134
+ "ece": 0.17639653883482279,
135
+ "sel_acc80": 0.9815789473684211,
136
+ "coverage": 1.0,
137
+ "false_abstain": 0.0,
138
+ "null_recall": "n/a",
139
+ "n": 475
140
+ }
141
+ },
142
+ "by_qtype": {},
143
+ "by_label_class": {
144
+ "answer": {
145
+ "nll": 0.6619388394682302,
146
+ "brier": 0.029335292543417674,
147
+ "jsd": 0.01347952047234738,
148
+ "acc": 0.9073684210526316,
149
+ "acc_k": 0.9073684210526316,
150
+ "ece": 0.17639653883482279,
151
+ "sel_acc80": 0.9815789473684211,
152
+ "coverage": 1.0,
153
+ "false_abstain": 0.0,
154
+ "null_recall": "n/a",
155
+ "n": 475
156
+ }
157
+ },
158
+ "baselines": {
159
+ "majority_class": 1,
160
+ "majority_acc": 0.5747368421052632
161
+ }
162
+ },
163
+ "wf_heldout_choice": {
164
+ "n": 1500,
165
+ "n_total": 1500,
166
+ "suffix_overflow": 0,
167
+ "raw": {
168
+ "nll": 0.08894784373441748,
169
+ "brier": 0.03804468692396801,
170
+ "jsd": 0.017131709690636845,
171
+ "acc": 0.978,
172
+ "acc_k": 0.9907997169143666,
173
+ "ece": 0.011830694854259511,
174
+ "sel_acc80": 0.9958333333333333,
175
+ "conf_wrong": 0.0,
176
+ "coverage": 0.9273333333333333,
177
+ "false_abstain": 0.01556970983722576,
178
+ "null_recall": 1.0,
179
+ "auroc_null": 0.9994834500654839,
180
+ "auroc_conf": 0.9995119213217171
181
+ },
182
+ "by_family": {
183
+ "tool_select": {
184
+ "nll": 0.08894784373441748,
185
+ "brier": 0.03804468692396801,
186
+ "jsd": 0.017131709690636845,
187
+ "acc": 0.978,
188
+ "acc_k": 0.9907997169143666,
189
+ "ece": 0.011830694854259511,
190
+ "sel_acc80": 0.9958333333333333,
191
+ "conf_wrong": 0.0,
192
+ "coverage": 0.9273333333333333,
193
+ "false_abstain": 0.01556970983722576,
194
+ "null_recall": 1.0,
195
+ "auroc_null": 0.9994834500654839,
196
+ "auroc_conf": 0.9995119213217171,
197
+ "n": 1500
198
+ }
199
+ },
200
+ "by_qtype": {
201
+ "choice": {
202
+ "nll": 0.08894784373441748,
203
+ "brier": 0.03804468692396801,
204
+ "jsd": 0.017131709690636845,
205
+ "acc": 0.978,
206
+ "acc_k": 0.9907997169143666,
207
+ "ece": 0.011830694854259511,
208
+ "sel_acc80": 0.9958333333333333,
209
+ "conf_wrong": 0.0,
210
+ "coverage": 0.9273333333333333,
211
+ "false_abstain": 0.01556970983722576,
212
+ "null_recall": 1.0,
213
+ "auroc_null": 0.9994834500654839,
214
+ "auroc_conf": 0.9995119213217171,
215
+ "n": 1500
216
+ }
217
+ },
218
+ "by_label_class": {
219
+ "answer": {
220
+ "nll": 0.09437487539006816,
221
+ "brier": 0.04038581321521866,
222
+ "jsd": 0.018169384833798215,
223
+ "acc": 0.9766454352441614,
224
+ "acc_k": 0.9907997169143666,
225
+ "ece": 0.012608277211354707,
226
+ "sel_acc80": 0.9946949602122016,
227
+ "coverage": 0.9844302901627743,
228
+ "false_abstain": 0.01556970983722576,
229
+ "null_recall": "n/a",
230
+ "n": 1413
231
+ },
232
+ "null": {
233
+ "nll": 0.0008053640857458489,
234
+ "brier": 2.1566814345354523e-05,
235
+ "jsd": 0.0002784340896366467,
236
+ "acc": 1.0,
237
+ "ece": 0.0007983151523546228,
238
+ "sel_acc80": 1.0,
239
+ "conf_wrong": 0.0,
240
+ "coverage": 0.0,
241
+ "false_abstain": "n/a",
242
+ "null_recall": 1.0,
243
+ "n": 87
244
+ }
245
+ },
246
+ "baselines": {
247
+ "majority_class": 2,
248
+ "majority_acc": 0.21333333333333335
249
+ }
250
+ },
251
+ "wf_heldout_noul": {
252
+ "n": 1500,
253
+ "n_total": 1500,
254
+ "suffix_overflow": 0,
255
+ "raw": {
256
+ "nll": 0.5731820141841979,
257
+ "brier": 0.30167717441890574,
258
+ "jsd": 0.10894462183976376,
259
+ "acc": 0.8106666666666666,
260
+ "acc_k": 0.8106666666666666,
261
+ "ece": 0.13052070760726933,
262
+ "sel_acc80": 0.88,
263
+ "coverage": 1.0,
264
+ "false_abstain": 0.0,
265
+ "null_recall": "n/a"
266
+ },
267
+ "by_family": {
268
+ "eligibility": {
269
+ "nll": 0.5731820141841979,
270
+ "brier": 0.30167717441890574,
271
+ "jsd": 0.10894462183976376,
272
+ "acc": 0.8106666666666666,
273
+ "acc_k": 0.8106666666666666,
274
+ "ece": 0.13052070760726933,
275
+ "sel_acc80": 0.88,
276
+ "coverage": 1.0,
277
+ "false_abstain": 0.0,
278
+ "null_recall": "n/a",
279
+ "n": 1500
280
+ }
281
+ },
282
+ "by_qtype": {
283
+ "noul": {
284
+ "nll": 0.5731820141841979,
285
+ "brier": 0.30167717441890574,
286
+ "jsd": 0.10894462183976376,
287
+ "acc": 0.8106666666666666,
288
+ "acc_k": 0.8106666666666666,
289
+ "ece": 0.13052070760726933,
290
+ "sel_acc80": 0.88,
291
+ "coverage": 1.0,
292
+ "false_abstain": 0.0,
293
+ "null_recall": "n/a",
294
+ "n": 1500
295
+ }
296
+ },
297
+ "by_label_class": {
298
+ "answer": {
299
+ "nll": 0.5731820141841979,
300
+ "brier": 0.30167717441890574,
301
+ "jsd": 0.10894462183976376,
302
+ "acc": 0.8106666666666666,
303
+ "acc_k": 0.8106666666666666,
304
+ "ece": 0.13052070760726933,
305
+ "sel_acc80": 0.88,
306
+ "coverage": 1.0,
307
+ "false_abstain": 0.0,
308
+ "null_recall": "n/a",
309
+ "n": 1500
310
+ }
311
+ },
312
+ "baselines": {
313
+ "majority_class": 0,
314
+ "majority_acc": 0.5833333333333334,
315
+ "always_positive_acc": 0.4166666666666667
316
+ }
317
+ },
318
+ "wf_heldout_score": {
319
+ "n": 1502,
320
+ "n_total": 1502,
321
+ "suffix_overflow": 0,
322
+ "raw": {
323
+ "nll": 1.0185536074805717,
324
+ "brier": 0.5644523897741041,
325
+ "jsd": 0.2801056596145076,
326
+ "acc": 0.5279627163781625,
327
+ "acc_k": 0.5279627163781625,
328
+ "ece": 0.05101245270667476,
329
+ "sel_acc80": 0.5806988352745425,
330
+ "coverage": 1.0,
331
+ "false_abstain": 0.0,
332
+ "null_recall": "n/a"
333
+ },
334
+ "by_family": {
335
+ "urgency": {
336
+ "nll": 1.0185536074805717,
337
+ "brier": 0.5644523897741041,
338
+ "jsd": 0.2801056596145076,
339
+ "acc": 0.5279627163781625,
340
+ "acc_k": 0.5279627163781625,
341
+ "ece": 0.05101245270667476,
342
+ "sel_acc80": 0.5806988352745425,
343
+ "coverage": 1.0,
344
+ "false_abstain": 0.0,
345
+ "null_recall": "n/a",
346
+ "n": 1502
347
+ }
348
+ },
349
+ "by_qtype": {
350
+ "score": {
351
+ "nll": 1.0185536074805717,
352
+ "brier": 0.5644523897741041,
353
+ "jsd": 0.2801056596145076,
354
+ "acc": 0.5279627163781625,
355
+ "acc_k": 0.5279627163781625,
356
+ "ece": 0.05101245270667476,
357
+ "sel_acc80": 0.5806988352745425,
358
+ "coverage": 1.0,
359
+ "false_abstain": 0.0,
360
+ "null_recall": "n/a",
361
+ "n": 1502
362
+ }
363
+ },
364
+ "by_label_class": {
365
+ "answer": {
366
+ "nll": 1.0185536074805717,
367
+ "brier": 0.5644523897741041,
368
+ "jsd": 0.2801056596145076,
369
+ "acc": 0.5279627163781625,
370
+ "acc_k": 0.5279627163781625,
371
+ "ece": 0.05101245270667476,
372
+ "sel_acc80": 0.5806988352745425,
373
+ "coverage": 1.0,
374
+ "false_abstain": 0.0,
375
+ "null_recall": "n/a",
376
+ "n": 1502
377
+ }
378
+ },
379
+ "baselines": {
380
+ "majority_class": 2,
381
+ "majority_acc": 0.3548601864181092
382
+ }
383
+ },
384
+ "wf_heldout_style": {
385
+ "n": 1627,
386
+ "n_total": 1627,
387
+ "suffix_overflow": 0,
388
+ "raw": {
389
+ "nll": 0.35412793927257424,
390
+ "brier": 0.2060959847410378,
391
+ "jsd": 0.10287023165278557,
392
+ "acc": 0.8586355255070682,
393
+ "acc_k": 0.8564294631710362,
394
+ "ece": 0.05414908969592812,
395
+ "sel_acc80": 0.9308755760368663,
396
+ "conf_wrong": 0.0,
397
+ "coverage": 0.9846342962507683,
398
+ "false_abstain": 0.0,
399
+ "null_recall": 1.0,
400
+ "auroc_null": 1.0,
401
+ "auroc_conf": 1.0
402
+ },
403
+ "by_family": {
404
+ "relevance": {
405
+ "nll": 0.6980276300431268,
406
+ "brier": 0.42102710856000347,
407
+ "jsd": 0.21389350686536507,
408
+ "acc": 0.72,
409
+ "acc_k": 0.72,
410
+ "ece": 0.17414700448513035,
411
+ "sel_acc80": 0.775,
412
+ "coverage": 1.0,
413
+ "false_abstain": 0.0,
414
+ "null_recall": "n/a",
415
+ "n": 150
416
+ },
417
+ "support": {
418
+ "nll": 0.20406603972307857,
419
+ "brier": 0.1248183124169987,
420
+ "jsd": 0.055781303712672675,
421
+ "acc": 0.912751677852349,
422
+ "acc_k": 0.912751677852349,
423
+ "ece": 0.043371342572589816,
424
+ "sel_acc80": 0.975,
425
+ "coverage": 1.0,
426
+ "false_abstain": 0.0,
427
+ "null_recall": "n/a",
428
+ "n": 149
429
+ },
430
+ "adequacy": {
431
+ "nll": 0.17072481903574357,
432
+ "brier": 0.10599008454380157,
433
+ "jsd": 0.047014351937127816,
434
+ "acc": 0.9266666666666666,
435
+ "acc_k": 0.9266666666666666,
436
+ "ece": 0.04129962166150417,
437
+ "sel_acc80": 0.9916666666666667,
438
+ "coverage": 1.0,
439
+ "false_abstain": 0.0,
440
+ "null_recall": "n/a",
441
+ "n": 150
442
+ },
443
+ "action_select": {
444
+ "nll": 0.231281707526007,
445
+ "brier": 0.13311344023592248,
446
+ "jsd": 0.058601355529894074,
447
+ "acc": 0.9133333333333333,
448
+ "acc_k": 0.9064748201438849,
449
+ "ece": 0.05428580959637955,
450
+ "sel_acc80": 0.9833333333333333,
451
+ "conf_wrong": 0.0,
452
+ "coverage": 0.9266666666666666,
453
+ "false_abstain": 0.0,
454
+ "null_recall": 1.0,
455
+ "auroc_null": 1.0,
456
+ "auroc_conf": 1.0,
457
+ "n": 150
458
+ },
459
+ "completeness": {
460
+ "nll": 0.7433181718309183,
461
+ "brier": 0.4258628433595575,
462
+ "jsd": 0.22361098694674933,
463
+ "acc": 0.7248322147651006,
464
+ "acc_k": 0.7248322147651006,
465
+ "ece": 0.19188565155803758,
466
+ "sel_acc80": 0.75,
467
+ "coverage": 1.0,
468
+ "false_abstain": 0.0,
469
+ "null_recall": "n/a",
470
+ "n": 149
471
+ },
472
+ "quality": {
473
+ "nll": 0.8425104976168987,
474
+ "brier": 0.4753130219919474,
475
+ "jsd": 0.24646774262839125,
476
+ "acc": 0.6259541984732825,
477
+ "acc_k": 0.6259541984732825,
478
+ "ece": 0.1351061500893294,
479
+ "sel_acc80": 0.6857142857142857,
480
+ "coverage": 1.0,
481
+ "false_abstain": 0.0,
482
+ "null_recall": "n/a",
483
+ "n": 131
484
+ },
485
+ "enum_extract": {
486
+ "nll": 0.0033332703444646995,
487
+ "brier": 0.00024124939750805901,
488
+ "jsd": 0.0011463143230046637,
489
+ "acc": 1.0,
490
+ "acc_k": 1.0,
491
+ "ece": 0.003245462179183957,
492
+ "sel_acc80": 1.0,
493
+ "conf_wrong": 0.0,
494
+ "coverage": 0.9666666666666667,
495
+ "false_abstain": 0.0,
496
+ "null_recall": 1.0,
497
+ "auroc_null": 1.0,
498
+ "auroc_conf": 1.0,
499
+ "n": 150
500
+ },
501
+ "categorical": {
502
+ "nll": 0.24690092856485552,
503
+ "brier": 0.14982400101731513,
504
+ "jsd": 0.07302109473186505,
505
+ "acc": 0.889763779527559,
506
+ "acc_k": 0.8870967741935484,
507
+ "ece": 0.053563406148294794,
508
+ "sel_acc80": 0.9607843137254902,
509
+ "conf_wrong": 0.0,
510
+ "coverage": 0.9763779527559056,
511
+ "false_abstain": 0.0,
512
+ "null_recall": 1.0,
513
+ "auroc_null": 1.0,
514
+ "auroc_conf": 1.0,
515
+ "n": 127
516
+ },
517
+ "routing": {
518
+ "nll": 0.21774816429206942,
519
+ "brier": 0.13021437367644959,
520
+ "jsd": 0.055718942276041364,
521
+ "acc": 0.9090909090909091,
522
+ "acc_k": 0.9047619047619048,
523
+ "ece": 0.05593117032990313,
524
+ "sel_acc80": 0.9905660377358491,
525
+ "conf_wrong": 0.0,
526
+ "coverage": 0.9545454545454546,
527
+ "false_abstain": 0.0,
528
+ "null_recall": 1.0,
529
+ "auroc_null": 1.0,
530
+ "auroc_conf": 1.0,
531
+ "n": 132
532
+ },
533
+ "severity": {
534
+ "nll": 0.6434271258602455,
535
+ "brier": 0.347468069379254,
536
+ "jsd": 0.19560876673722952,
537
+ "acc": 0.7974683544303798,
538
+ "acc_k": 0.7974683544303798,
539
+ "ece": 0.22888638286650936,
540
+ "sel_acc80": 0.875,
541
+ "coverage": 1.0,
542
+ "false_abstain": 0.0,
543
+ "null_recall": "n/a",
544
+ "n": 79
545
+ },
546
+ "policy_permit": {
547
+ "nll": 0.08175127275497936,
548
+ "brier": 0.03988201939839468,
549
+ "jsd": 0.02053868180510852,
550
+ "acc": 0.9798657718120806,
551
+ "acc_k": 0.9798657718120806,
552
+ "ece": 0.031195846179987764,
553
+ "sel_acc80": 1.0,
554
+ "coverage": 1.0,
555
+ "false_abstain": 0.0,
556
+ "null_recall": "n/a",
557
+ "n": 149
558
+ },
559
+ "fact": {
560
+ "nll": 0.3245140742970864,
561
+ "brier": 0.201230115726297,
562
+ "jsd": 0.09200789581633895,
563
+ "acc": 0.8558558558558559,
564
+ "acc_k": 0.8558558558558559,
565
+ "ece": 0.0778693476238766,
566
+ "sel_acc80": 0.9213483146067416,
567
+ "coverage": 1.0,
568
+ "false_abstain": 0.0,
569
+ "null_recall": "n/a",
570
+ "n": 111
571
+ }
572
+ },
573
+ "by_qtype": {
574
+ "score": {
575
+ "nll": 0.7399964051867366,
576
+ "brier": 0.4249972756512392,
577
+ "jsd": 0.2222837523407285,
578
+ "acc": 0.7092337917485265,
579
+ "acc_k": 0.7092337917485265,
580
+ "ece": 0.1720800453354193,
581
+ "sel_acc80": 0.7549019607843137,
582
+ "coverage": 1.0,
583
+ "false_abstain": 0.0,
584
+ "null_recall": "n/a",
585
+ "n": 509
586
+ },
587
+ "noul": {
588
+ "nll": 0.186433926049318,
589
+ "brier": 0.11229947221410162,
590
+ "jsd": 0.0512284562939753,
591
+ "acc": 0.9230769230769231,
592
+ "acc_k": 0.9230769230769231,
593
+ "ece": 0.024164358063972435,
594
+ "sel_acc80": 0.9754464285714286,
595
+ "coverage": 1.0,
596
+ "false_abstain": 0.0,
597
+ "null_recall": "n/a",
598
+ "n": 559
599
+ },
600
+ "choice": {
601
+ "nll": 0.1704676606348132,
602
+ "brier": 0.10057092826387289,
603
+ "jsd": 0.04577948101845999,
604
+ "acc": 0.9302325581395349,
605
+ "acc_k": 0.9269662921348315,
606
+ "ece": 0.0223422101657164,
607
+ "sel_acc80": 0.9933035714285714,
608
+ "conf_wrong": 0.0,
609
+ "coverage": 0.9552772808586762,
610
+ "false_abstain": 0.0,
611
+ "null_recall": 1.0,
612
+ "auroc_null": 1.0,
613
+ "auroc_conf": 1.0,
614
+ "n": 559
615
+ }
616
+ },
617
+ "by_label_class": {
618
+ "answer": {
619
+ "nll": 0.35959165383560376,
620
+ "brier": 0.2093046663699308,
621
+ "jsd": 0.10445406727566603,
622
+ "acc": 0.8564294631710362,
623
+ "acc_k": 0.8564294631710362,
624
+ "ece": 0.05493348601568356,
625
+ "sel_acc80": 0.9297971918876755,
626
+ "coverage": 1.0,
627
+ "false_abstain": 0.0,
628
+ "null_recall": "n/a",
629
+ "n": 1602
630
+ },
631
+ "null": {
632
+ "nll": 0.004013110073643866,
633
+ "brier": 0.0004836659615737134,
634
+ "jsd": 0.0013780449386066767,
635
+ "acc": 1.0,
636
+ "ece": 0.003884973526001012,
637
+ "sel_acc80": 1.0,
638
+ "conf_wrong": 0.0,
639
+ "coverage": 0.0,
640
+ "false_abstain": "n/a",
641
+ "null_recall": 1.0,
642
+ "n": 25
643
+ }
644
+ },
645
+ "baselines": {
646
+ "majority_class": 1,
647
+ "majority_acc": 0.36447449293177625
648
+ }
649
+ },
650
+ "wf_rubric_flip": {
651
+ "n": 2062,
652
+ "n_total": 2062,
653
+ "suffix_overflow": 0,
654
+ "raw": {
655
+ "nll": 0.5140816127009797,
656
+ "brier": 0.27501173451171385,
657
+ "jsd": 0.1274834860045779,
658
+ "acc": 0.7856450048496605,
659
+ "acc_k": 0.7914645974781765,
660
+ "ece": 0.04033603424533379,
661
+ "sel_acc80": 0.8878787878787879,
662
+ "coverage": 0.9932104752667313,
663
+ "false_abstain": 0.006789524733268671,
664
+ "null_recall": "n/a"
665
+ },
666
+ "by_family": {
667
+ "urgency": {
668
+ "nll": 1.0503140999721947,
669
+ "brier": 0.5826411232348119,
670
+ "jsd": 0.2870499222013987,
671
+ "acc": 0.5142857142857142,
672
+ "acc_k": 0.5142857142857142,
673
+ "ece": 0.061800167475427895,
674
+ "sel_acc80": 0.5464285714285714,
675
+ "coverage": 1.0,
676
+ "false_abstain": 0.0,
677
+ "null_recall": "n/a",
678
+ "n": 700
679
+ },
680
+ "eligibility": {
681
+ "nll": 0.3789669930673108,
682
+ "brier": 0.19723455708945753,
683
+ "jsd": 0.07565788618043696,
684
+ "acc": 0.8704268292682927,
685
+ "acc_k": 0.8704268292682927,
686
+ "ece": 0.0704190259481349,
687
+ "sel_acc80": 0.9314285714285714,
688
+ "coverage": 1.0,
689
+ "false_abstain": 0.0,
690
+ "null_recall": "n/a",
691
+ "n": 656
692
+ },
693
+ "tool_select": {
694
+ "nll": 0.10795193761576238,
695
+ "brier": 0.042265638595044515,
696
+ "jsd": 0.017428370065288763,
697
+ "acc": 0.9759206798866855,
698
+ "acc_k": 0.9929178470254958,
699
+ "ece": 0.015375537465381884,
700
+ "sel_acc80": 0.9946902654867257,
701
+ "coverage": 0.9801699716713881,
702
+ "false_abstain": 0.019830028328611898,
703
+ "null_recall": "n/a",
704
+ "n": 706
705
+ }
706
+ },
707
+ "by_qtype": {
708
+ "score": {
709
+ "nll": 1.0503140999721947,
710
+ "brier": 0.5826411232348119,
711
+ "jsd": 0.2870499222013987,
712
+ "acc": 0.5142857142857142,
713
+ "acc_k": 0.5142857142857142,
714
+ "ece": 0.061800167475427895,
715
+ "sel_acc80": 0.5464285714285714,
716
+ "coverage": 1.0,
717
+ "false_abstain": 0.0,
718
+ "null_recall": "n/a",
719
+ "n": 700
720
+ },
721
+ "noul": {
722
+ "nll": 0.3789669930673108,
723
+ "brier": 0.19723455708945753,
724
+ "jsd": 0.07565788618043696,
725
+ "acc": 0.8704268292682927,
726
+ "acc_k": 0.8704268292682927,
727
+ "ece": 0.0704190259481349,
728
+ "sel_acc80": 0.9314285714285714,
729
+ "coverage": 1.0,
730
+ "false_abstain": 0.0,
731
+ "null_recall": "n/a",
732
+ "n": 656
733
+ },
734
+ "choice": {
735
+ "nll": 0.10795193761576238,
736
+ "brier": 0.042265638595044515,
737
+ "jsd": 0.017428370065288763,
738
+ "acc": 0.9759206798866855,
739
+ "acc_k": 0.9929178470254958,
740
+ "ece": 0.015375537465381884,
741
+ "sel_acc80": 0.9946902654867257,
742
+ "coverage": 0.9801699716713881,
743
+ "false_abstain": 0.019830028328611898,
744
+ "null_recall": "n/a",
745
+ "n": 706
746
+ }
747
+ },
748
+ "by_label_class": {
749
+ "answer": {
750
+ "nll": 0.5140816127009797,
751
+ "brier": 0.27501173451171385,
752
+ "jsd": 0.1274834860045779,
753
+ "acc": 0.7856450048496605,
754
+ "acc_k": 0.7914645974781765,
755
+ "ece": 0.04033603424533379,
756
+ "sel_acc80": 0.8878787878787879,
757
+ "coverage": 0.9932104752667313,
758
+ "false_abstain": 0.006789524733268671,
759
+ "null_recall": "n/a",
760
+ "n": 2062
761
+ }
762
+ },
763
+ "baselines": {
764
+ "majority_class": 0,
765
+ "majority_acc": 0.32104752667313285
766
+ },
767
+ "flip": {
768
+ "n_pairs": 1031,
769
+ "flip_rate": 0.8205625606207565,
770
+ "both_correct": 0.6702230843840931,
771
+ "acc": 0.7914645974781765
772
+ },
773
+ "flip_by_family": {
774
+ "eligibility": {
775
+ "n_pairs": 328,
776
+ "flip_rate": 0.7591463414634146,
777
+ "both_correct": 0.75,
778
+ "acc": 0.8704268292682927
779
+ },
780
+ "tool_select": {
781
+ "n_pairs": 353,
782
+ "flip_rate": 0.9915014164305949,
783
+ "both_correct": 0.9858356940509915,
784
+ "acc": 0.9929178470254958
785
+ },
786
+ "urgency": {
787
+ "n_pairs": 350,
788
+ "flip_rate": 0.7057142857142857,
789
+ "both_correct": 0.27714285714285714,
790
+ "acc": 0.5142857142857142
791
+ }
792
+ }
793
+ },
794
+ "wf_rubric_shuffled": {
795
+ "n": 3775,
796
+ "n_total": 3775,
797
+ "suffix_overflow": 0,
798
+ "raw": {
799
+ "nll": 3.3019875953051896,
800
+ "brier": 0.9847791419883737,
801
+ "jsd": 0.38205552797027176,
802
+ "acc": 0.3801324503311258,
803
+ "acc_k": 0.3811920529801324,
804
+ "ece": 0.39123494437198764,
805
+ "sel_acc80": 0.39470198675496687,
806
+ "coverage": 0.9925827814569537,
807
+ "false_abstain": 0.007417218543046358,
808
+ "null_recall": "n/a"
809
+ },
810
+ "by_family": {
811
+ "tool_select": {
812
+ "nll": 8.01855352489357,
813
+ "brier": 1.585383500778014,
814
+ "jsd": 0.5686605163566438,
815
+ "acc": 0.15359116022099448,
816
+ "acc_k": 0.1580110497237569,
817
+ "ece": 0.7718642068172686,
818
+ "sel_acc80": 0.13535911602209943,
819
+ "coverage": 0.969060773480663,
820
+ "false_abstain": 0.030939226519337018,
821
+ "null_recall": "n/a",
822
+ "n": 905
823
+ },
824
+ "eligibility": {
825
+ "nll": 2.0992631849983714,
826
+ "brier": 0.8251410346340735,
827
+ "jsd": 0.28859514090273525,
828
+ "acc": 0.5234758234057463,
829
+ "acc_k": 0.5234758234057463,
830
+ "ece": 0.39013672221684337,
831
+ "sel_acc80": 0.5332749562171629,
832
+ "coverage": 1.0,
833
+ "false_abstain": 0.0,
834
+ "null_recall": "n/a",
835
+ "n": 1427
836
+ },
837
+ "urgency": {
838
+ "nll": 1.5333081547163776,
839
+ "brier": 0.7659687708795461,
840
+ "jsd": 0.35744739065613984,
841
+ "acc": 0.3804573804573805,
842
+ "acc_k": 0.3804573804573805,
843
+ "ece": 0.15649522523952702,
844
+ "sel_acc80": 0.38961038961038963,
845
+ "coverage": 1.0,
846
+ "false_abstain": 0.0,
847
+ "null_recall": "n/a",
848
+ "n": 1443
849
+ }
850
+ },
851
+ "by_qtype": {
852
+ "choice": {
853
+ "nll": 8.01855352489357,
854
+ "brier": 1.585383500778014,
855
+ "jsd": 0.5686605163566438,
856
+ "acc": 0.15359116022099448,
857
+ "acc_k": 0.1580110497237569,
858
+ "ece": 0.7718642068172686,
859
+ "sel_acc80": 0.13535911602209943,
860
+ "coverage": 0.969060773480663,
861
+ "false_abstain": 0.030939226519337018,
862
+ "null_recall": "n/a",
863
+ "n": 905
864
+ },
865
+ "noul": {
866
+ "nll": 2.0992631849983714,
867
+ "brier": 0.8251410346340735,
868
+ "jsd": 0.28859514090273525,
869
+ "acc": 0.5234758234057463,
870
+ "acc_k": 0.5234758234057463,
871
+ "ece": 0.39013672221684337,
872
+ "sel_acc80": 0.5332749562171629,
873
+ "coverage": 1.0,
874
+ "false_abstain": 0.0,
875
+ "null_recall": "n/a",
876
+ "n": 1427
877
+ },
878
+ "score": {
879
+ "nll": 1.5333081547163776,
880
+ "brier": 0.7659687708795461,
881
+ "jsd": 0.35744739065613984,
882
+ "acc": 0.3804573804573805,
883
+ "acc_k": 0.3804573804573805,
884
+ "ece": 0.15649522523952702,
885
+ "sel_acc80": 0.38961038961038963,
886
+ "coverage": 1.0,
887
+ "false_abstain": 0.0,
888
+ "null_recall": "n/a",
889
+ "n": 1443
890
+ }
891
+ },
892
+ "by_label_class": {
893
+ "answer": {
894
+ "nll": 3.3019875953051896,
895
+ "brier": 0.9847791419883737,
896
+ "jsd": 0.38205552797027176,
897
+ "acc": 0.3801324503311258,
898
+ "acc_k": 0.3811920529801324,
899
+ "ece": 0.39123494437198764,
900
+ "sel_acc80": 0.39470198675496687,
901
+ "coverage": 0.9925827814569537,
902
+ "false_abstain": 0.007417218543046358,
903
+ "null_recall": "n/a",
904
+ "n": 3775
905
+ }
906
+ },
907
+ "baselines": {
908
+ "majority_class": 0,
909
+ "majority_acc": 0.36
910
+ }
911
+ },
912
+ "cua_s1_forms_demo": {
913
+ "n": 196,
914
+ "n_total": 196,
915
+ "suffix_overflow": 0,
916
+ "raw": {
917
+ "nll": 0.0006697119946596961,
918
+ "brier": 2.5804204528327477e-06,
919
+ "jsd": 0.00023202367765799874,
920
+ "acc": 1.0,
921
+ "acc_k": 1.0,
922
+ "ece": 0.0006688650773496452,
923
+ "sel_acc80": 1.0,
924
+ "coverage": 1.0,
925
+ "false_abstain": 0.0,
926
+ "null_recall": "n/a"
927
+ },
928
+ "by_family": {
929
+ "cua_s1_forms": {
930
+ "nll": 0.0006697119946596961,
931
+ "brier": 2.5804204528327477e-06,
932
+ "jsd": 0.00023202367765799874,
933
+ "acc": 1.0,
934
+ "acc_k": 1.0,
935
+ "ece": 0.0006688650773496452,
936
+ "sel_acc80": 1.0,
937
+ "coverage": 1.0,
938
+ "false_abstain": 0.0,
939
+ "null_recall": "n/a",
940
+ "n": 196
941
+ }
942
+ },
943
+ "by_qtype": {
944
+ "choice": {
945
+ "nll": 0.0006697119946596961,
946
+ "brier": 2.5804204528327477e-06,
947
+ "jsd": 0.00023202367765799874,
948
+ "acc": 1.0,
949
+ "acc_k": 1.0,
950
+ "ece": 0.0006688650773496452,
951
+ "sel_acc80": 1.0,
952
+ "coverage": 1.0,
953
+ "false_abstain": 0.0,
954
+ "null_recall": "n/a",
955
+ "n": 196
956
+ }
957
+ },
958
+ "by_label_class": {
959
+ "answer": {
960
+ "nll": 0.0006697119946596961,
961
+ "brier": 2.5804204528327477e-06,
962
+ "jsd": 0.00023202367765799874,
963
+ "acc": 1.0,
964
+ "acc_k": 1.0,
965
+ "ece": 0.0006688650773496452,
966
+ "sel_acc80": 1.0,
967
+ "coverage": 1.0,
968
+ "false_abstain": 0.0,
969
+ "null_recall": "n/a",
970
+ "n": 196
971
+ }
972
+ },
973
+ "baselines": {
974
+ "majority_class": 18,
975
+ "majority_acc": 0.27040816326530615
976
+ }
977
+ },
978
+ "cua_s1_forms_test": {
979
+ "n": 24370,
980
+ "n_total": 24370,
981
+ "suffix_overflow": 0,
982
+ "raw": {
983
+ "nll": 0.032467243628837364,
984
+ "brier": 0.01572971627781683,
985
+ "jsd": 0.006402877388207127,
986
+ "acc": 0.9902749281903981,
987
+ "acc_k": 0.9902749281903981,
988
+ "ece": 0.004090282016471624,
989
+ "sel_acc80": 0.9998974148543291,
990
+ "coverage": 1.0,
991
+ "false_abstain": 0.0,
992
+ "null_recall": "n/a"
993
+ },
994
+ "by_family": {
995
+ "cua_s1_forms": {
996
+ "nll": 0.032467243628837364,
997
+ "brier": 0.01572971627781683,
998
+ "jsd": 0.006402877388207127,
999
+ "acc": 0.9902749281903981,
1000
+ "acc_k": 0.9902749281903981,
1001
+ "ece": 0.004090282016471624,
1002
+ "sel_acc80": 0.9998974148543291,
1003
+ "coverage": 1.0,
1004
+ "false_abstain": 0.0,
1005
+ "null_recall": "n/a",
1006
+ "n": 24370
1007
+ }
1008
+ },
1009
+ "by_qtype": {
1010
+ "choice": {
1011
+ "nll": 0.032467243628837364,
1012
+ "brier": 0.01572971627781683,
1013
+ "jsd": 0.006402877388207127,
1014
+ "acc": 0.9902749281903981,
1015
+ "acc_k": 0.9902749281903981,
1016
+ "ece": 0.004090282016471624,
1017
+ "sel_acc80": 0.9998974148543291,
1018
+ "coverage": 1.0,
1019
+ "false_abstain": 0.0,
1020
+ "null_recall": "n/a",
1021
+ "n": 24370
1022
+ }
1023
+ },
1024
+ "by_label_class": {
1025
+ "answer": {
1026
+ "nll": 0.032467243628837364,
1027
+ "brier": 0.01572971627781683,
1028
+ "jsd": 0.006402877388207127,
1029
+ "acc": 0.9902749281903981,
1030
+ "acc_k": 0.9902749281903981,
1031
+ "ece": 0.004090282016471624,
1032
+ "sel_acc80": 0.9998974148543291,
1033
+ "coverage": 1.0,
1034
+ "false_abstain": 0.0,
1035
+ "null_recall": "n/a",
1036
+ "n": 24370
1037
+ }
1038
+ },
1039
+ "baselines": {
1040
+ "majority_class": 16,
1041
+ "majority_acc": 0.07033237587197373
1042
+ }
1043
+ },
1044
+ "jev4b_adversarial": {
1045
+ "n": 600,
1046
+ "n_total": 600,
1047
+ "suffix_overflow": 0,
1048
+ "raw": {
1049
+ "nll": 0.21596318335730846,
1050
+ "brier": 0.11449747234656535,
1051
+ "jsd": 0.0566693510735706,
1052
+ "acc": 0.9266666666666666,
1053
+ "acc_k": 0.9266666666666666,
1054
+ "ece": 0.033218660900990224,
1055
+ "sel_acc80": 0.98125,
1056
+ "coverage": 1.0,
1057
+ "false_abstain": 0.0,
1058
+ "null_recall": "n/a"
1059
+ },
1060
+ "by_family": {
1061
+ "jev4b_distill": {
1062
+ "nll": 0.21596318335730846,
1063
+ "brier": 0.11449747234656535,
1064
+ "jsd": 0.0566693510735706,
1065
+ "acc": 0.9266666666666666,
1066
+ "acc_k": 0.9266666666666666,
1067
+ "ece": 0.033218660900990224,
1068
+ "sel_acc80": 0.98125,
1069
+ "coverage": 1.0,
1070
+ "false_abstain": 0.0,
1071
+ "null_recall": "n/a",
1072
+ "n": 600
1073
+ }
1074
+ },
1075
+ "by_qtype": {
1076
+ "choice": {
1077
+ "nll": 0.15764157777699858,
1078
+ "brier": 0.07730203604858563,
1079
+ "jsd": 0.03353047215992909,
1080
+ "acc": 0.9527777777777777,
1081
+ "acc_k": 0.9527777777777777,
1082
+ "ece": 0.027492753830220998,
1083
+ "sel_acc80": 0.9895833333333334,
1084
+ "coverage": 1.0,
1085
+ "false_abstain": 0.0,
1086
+ "null_recall": "n/a",
1087
+ "n": 360
1088
+ },
1089
+ "score": {
1090
+ "nll": 0.588537222505538,
1091
+ "brier": 0.3360254493815728,
1092
+ "jsd": 0.17654749465911643,
1093
+ "acc": 0.775,
1094
+ "acc_k": 0.775,
1095
+ "ece": 0.11445025528470672,
1096
+ "sel_acc80": 0.8125,
1097
+ "coverage": 1.0,
1098
+ "false_abstain": 0.0,
1099
+ "null_recall": "n/a",
1100
+ "n": 120
1101
+ },
1102
+ "noul": {
1103
+ "nll": 0.018353960950008512,
1104
+ "brier": 0.004555804205497079,
1105
+ "jsd": 0.006207844228949319,
1106
+ "acc": 1.0,
1107
+ "acc_k": 1.0,
1108
+ "ece": 0.01696336468060807,
1109
+ "sel_acc80": 1.0,
1110
+ "coverage": 1.0,
1111
+ "false_abstain": 0.0,
1112
+ "null_recall": "n/a",
1113
+ "n": 120
1114
+ }
1115
+ },
1116
+ "by_label_class": {
1117
+ "answer": {
1118
+ "nll": 0.21596318335730846,
1119
+ "brier": 0.11449747234656535,
1120
+ "jsd": 0.0566693510735706,
1121
+ "acc": 0.9266666666666666,
1122
+ "acc_k": 0.9266666666666666,
1123
+ "ece": 0.033218660900990224,
1124
+ "sel_acc80": 0.98125,
1125
+ "coverage": 1.0,
1126
+ "false_abstain": 0.0,
1127
+ "null_recall": "n/a",
1128
+ "n": 600
1129
+ }
1130
+ },
1131
+ "baselines": {
1132
+ "majority_class": 0,
1133
+ "majority_acc": 0.42833333333333334
1134
+ }
1135
+ },
1136
+ "jevlogs_triage": {
1137
+ "n": 5080,
1138
+ "n_total": 5080,
1139
+ "suffix_overflow": 0,
1140
+ "raw": {
1141
+ "nll": 0.8200663549047869,
1142
+ "brier": 0.5245085358820047,
1143
+ "jsd": 0.19936215850961117,
1144
+ "acc": 0.6734251968503937,
1145
+ "acc_k": 0.6734251968503937,
1146
+ "ece": 0.19179721678805164,
1147
+ "sel_acc80": 0.6801181102362205,
1148
+ "coverage": 1.0,
1149
+ "false_abstain": 0.0,
1150
+ "null_recall": "n/a"
1151
+ },
1152
+ "by_family": {
1153
+ "log_hdfs": {
1154
+ "nll": 0.8322974851760632,
1155
+ "brier": 0.5211627531136993,
1156
+ "jsd": 0.1981452827621227,
1157
+ "acc": 0.6664,
1158
+ "acc_k": 0.6664,
1159
+ "ece": 0.17615253834724426,
1160
+ "sel_acc80": 0.6785,
1161
+ "coverage": 1.0,
1162
+ "false_abstain": 0.0,
1163
+ "null_recall": "n/a",
1164
+ "n": 2500
1165
+ },
1166
+ "log_bgl": {
1167
+ "nll": 0.7998629474253376,
1168
+ "brier": 0.521462450412092,
1169
+ "jsd": 0.1985896785858705,
1170
+ "acc": 0.6856,
1171
+ "acc_k": 0.6856,
1172
+ "ece": 0.24997946472167967,
1173
+ "sel_acc80": 0.659,
1174
+ "coverage": 1.0,
1175
+ "false_abstain": 0.0,
1176
+ "null_recall": "n/a",
1177
+ "n": 2500
1178
+ },
1179
+ "log_adversarial": {
1180
+ "nll": 1.0692000176602048,
1181
+ "brier": 0.7242544183263233,
1182
+ "jsd": 0.26152952323552203,
1183
+ "acc": 0.5125,
1184
+ "acc_k": 0.5125,
1185
+ "ece": 0.304364737868309,
1186
+ "sel_acc80": 0.515625,
1187
+ "coverage": 1.0,
1188
+ "false_abstain": 0.0,
1189
+ "null_recall": "n/a",
1190
+ "n": 80
1191
+ }
1192
+ },
1193
+ "by_qtype": {
1194
+ "noul": {
1195
+ "nll": 0.8200663549047869,
1196
+ "brier": 0.5245085358820047,
1197
+ "jsd": 0.19936215850961117,
1198
+ "acc": 0.6734251968503937,
1199
+ "acc_k": 0.6734251968503937,
1200
+ "ece": 0.19179721678805164,
1201
+ "sel_acc80": 0.6801181102362205,
1202
+ "coverage": 1.0,
1203
+ "false_abstain": 0.0,
1204
+ "null_recall": "n/a",
1205
+ "n": 5080
1206
+ }
1207
+ },
1208
+ "by_label_class": {
1209
+ "answer": {
1210
+ "nll": 0.8200663549047869,
1211
+ "brier": 0.5245085358820047,
1212
+ "jsd": 0.19936215850961117,
1213
+ "acc": 0.6734251968503937,
1214
+ "acc_k": 0.6734251968503937,
1215
+ "ece": 0.19179721678805164,
1216
+ "sel_acc80": 0.6801181102362205,
1217
+ "coverage": 1.0,
1218
+ "false_abstain": 0.0,
1219
+ "null_recall": "n/a",
1220
+ "n": 5080
1221
+ }
1222
+ },
1223
+ "baselines": {
1224
+ "majority_class": 0,
1225
+ "majority_acc": 0.6968503937007874,
1226
+ "always_positive_acc": 0.3031496062992126
1227
+ }
1228
+ },
1229
+ "mind2web_choice": {
1230
+ "n": 1600,
1231
+ "n_total": 1600,
1232
+ "suffix_overflow": 0,
1233
+ "raw": {
1234
+ "nll": 1.2406711424343253,
1235
+ "brier": 0.5732404237664536,
1236
+ "jsd": 0.2788736075063562,
1237
+ "acc": 0.544375,
1238
+ "acc_k": 0.544375,
1239
+ "ece": 0.08040638350881635,
1240
+ "sel_acc80": 0.615625,
1241
+ "coverage": 1.0,
1242
+ "false_abstain": 0.0,
1243
+ "null_recall": "n/a"
1244
+ },
1245
+ "by_family": {
1246
+ "mind2web": {
1247
+ "nll": 1.2406711424343253,
1248
+ "brier": 0.5732404237664536,
1249
+ "jsd": 0.2788736075063562,
1250
+ "acc": 0.544375,
1251
+ "acc_k": 0.544375,
1252
+ "ece": 0.08040638350881635,
1253
+ "sel_acc80": 0.615625,
1254
+ "coverage": 1.0,
1255
+ "false_abstain": 0.0,
1256
+ "null_recall": "n/a",
1257
+ "n": 1600
1258
+ }
1259
+ },
1260
+ "by_qtype": {
1261
+ "choice": {
1262
+ "nll": 1.8521192870863956,
1263
+ "brier": 0.7087342212760166,
1264
+ "jsd": 0.3756742847617898,
1265
+ "acc": 0.44,
1266
+ "acc_k": 0.44,
1267
+ "ece": 0.08787944102659824,
1268
+ "sel_acc80": 0.49375,
1269
+ "coverage": 1.0,
1270
+ "false_abstain": 0.0,
1271
+ "null_recall": "n/a",
1272
+ "n": 800
1273
+ },
1274
+ "noul": {
1275
+ "nll": 0.6292229977822548,
1276
+ "brier": 0.4377466262568903,
1277
+ "jsd": 0.18207293025092264,
1278
+ "acc": 0.64875,
1279
+ "acc_k": 0.64875,
1280
+ "ece": 0.07293332599103454,
1281
+ "sel_acc80": 0.6875,
1282
+ "coverage": 1.0,
1283
+ "false_abstain": 0.0,
1284
+ "null_recall": "n/a",
1285
+ "n": 800
1286
+ }
1287
+ },
1288
+ "by_label_class": {
1289
+ "answer": {
1290
+ "nll": 1.2406711424343253,
1291
+ "brier": 0.5732404237664536,
1292
+ "jsd": 0.2788736075063562,
1293
+ "acc": 0.544375,
1294
+ "acc_k": 0.544375,
1295
+ "ece": 0.08040638350881635,
1296
+ "sel_acc80": 0.615625,
1297
+ "coverage": 1.0,
1298
+ "false_abstain": 0.0,
1299
+ "null_recall": "n/a",
1300
+ "n": 1600
1301
+ }
1302
+ },
1303
+ "baselines": {
1304
+ "majority_class": 0,
1305
+ "majority_acc": 0.426875
1306
+ }
1307
+ },
1308
+ "pagerduty_trigger": {
1309
+ "n": 6000,
1310
+ "n_total": 6000,
1311
+ "suffix_overflow": 0,
1312
+ "raw": {
1313
+ "nll": 0.396707458891549,
1314
+ "brier": 0.23002239391086174,
1315
+ "jsd": 0.11322283340814882,
1316
+ "acc": 0.838,
1317
+ "acc_k": 0.838,
1318
+ "ece": 0.07345882829030351,
1319
+ "sel_acc80": 0.8979166666666667,
1320
+ "coverage": 1.0,
1321
+ "false_abstain": 0.0,
1322
+ "null_recall": "n/a"
1323
+ },
1324
+ "by_family": {
1325
+ "replica_lag_data_risk": {
1326
+ "nll": 0.29740572278535377,
1327
+ "brier": 0.14798980251485705,
1328
+ "jsd": 0.09449495177807844,
1329
+ "acc": 1.0,
1330
+ "acc_k": 1.0,
1331
+ "ece": 0.2269586962566041,
1332
+ "sel_acc80": 1.0,
1333
+ "coverage": 1.0,
1334
+ "false_abstain": 0.0,
1335
+ "null_recall": "n/a",
1336
+ "n": 114
1337
+ },
1338
+ "tls_hours_left_serving": {
1339
+ "nll": 2.462503098332096,
1340
+ "brier": 1.4718128845820069,
1341
+ "jsd": 0.5430561946798185,
1342
+ "acc": 0.0,
1343
+ "acc_k": 0.0,
1344
+ "ece": 0.7213651537895203,
1345
+ "sel_acc80": 0.0,
1346
+ "coverage": 1.0,
1347
+ "false_abstain": 0.0,
1348
+ "null_recall": "n/a",
1349
+ "n": 98
1350
+ },
1351
+ "disk_full_writes_failing": {
1352
+ "nll": 0.6112621743776719,
1353
+ "brier": 0.43961606766122513,
1354
+ "jsd": 0.1828291039570283,
1355
+ "acc": 0.5,
1356
+ "acc_k": 0.5,
1357
+ "ece": 0.40561820756305345,
1358
+ "sel_acc80": 0.625,
1359
+ "coverage": 1.0,
1360
+ "false_abstain": 0.0,
1361
+ "null_recall": "n/a",
1362
+ "n": 110
1363
+ },
1364
+ "revoked_key_used": {
1365
+ "nll": 2.25962306280277,
1366
+ "brier": 1.3607013127715883,
1367
+ "jsd": 0.49743537591946874,
1368
+ "acc": 0.0,
1369
+ "acc_k": 0.0,
1370
+ "ece": 0.7095300555229187,
1371
+ "sel_acc80": 0.0,
1372
+ "coverage": 1.0,
1373
+ "false_abstain": 0.0,
1374
+ "null_recall": "n/a",
1375
+ "n": 102
1376
+ },
1377
+ "checksum_corruption": {
1378
+ "nll": 0.16655108974517663,
1379
+ "brier": 0.05001836994744468,
1380
+ "jsd": 0.05564429813812224,
1381
+ "acc": 1.0,
1382
+ "acc_k": 1.0,
1383
+ "ece": 0.14739903873867463,
1384
+ "sel_acc80": 1.0,
1385
+ "coverage": 1.0,
1386
+ "false_abstain": 0.0,
1387
+ "null_recall": "n/a",
1388
+ "n": 90
1389
+ },
1390
+ "checkout_slo_burn": {
1391
+ "nll": 0.8234049439981885,
1392
+ "brier": 0.5886945952552634,
1393
+ "jsd": 0.24262392305224867,
1394
+ "acc": 0.5,
1395
+ "acc_k": 0.5,
1396
+ "ece": 0.5280300080776215,
1397
+ "sel_acc80": 0.375,
1398
+ "coverage": 1.0,
1399
+ "false_abstain": 0.0,
1400
+ "null_recall": "n/a",
1401
+ "n": 100
1402
+ },
1403
+ "payments_processor_down": {
1404
+ "nll": 0.7458985031769221,
1405
+ "brier": 0.5397173510738811,
1406
+ "jsd": 0.21208411419957937,
1407
+ "acc": 0.5,
1408
+ "acc_k": 0.5,
1409
+ "ece": 0.3847963486580139,
1410
+ "sel_acc80": 0.618421052631579,
1411
+ "coverage": 1.0,
1412
+ "false_abstain": 0.0,
1413
+ "null_recall": "n/a",
1414
+ "n": 94
1415
+ },
1416
+ "crash_loop_user_impact": {
1417
+ "nll": 0.45723675404949216,
1418
+ "brier": 0.28838964871857825,
1419
+ "jsd": 0.14411361556383404,
1420
+ "acc": 0.7916666666666666,
1421
+ "acc_k": 0.7916666666666666,
1422
+ "ece": 0.3439861076573531,
1423
+ "sel_acc80": 0.7402597402597403,
1424
+ "coverage": 1.0,
1425
+ "false_abstain": 0.0,
1426
+ "null_recall": "n/a",
1427
+ "n": 96
1428
+ },
1429
+ "credential_stuffing": {
1430
+ "nll": 0.6855976842756064,
1431
+ "brier": 0.5141043463970237,
1432
+ "jsd": 0.19799073218670762,
1433
+ "acc": 0.5,
1434
+ "acc_k": 0.5,
1435
+ "ece": 0.4205113059530656,
1436
+ "sel_acc80": 0.6233766233766234,
1437
+ "coverage": 1.0,
1438
+ "false_abstain": 0.0,
1439
+ "null_recall": "n/a",
1440
+ "n": 96
1441
+ },
1442
+ "queue_backup_checkout": {
1443
+ "nll": 0.4943918081315674,
1444
+ "brier": 0.3165144497680829,
1445
+ "jsd": 0.1554112283610238,
1446
+ "acc": 0.5,
1447
+ "acc_k": 0.5,
1448
+ "ece": 0.3690876066684723,
1449
+ "sel_acc80": 0.625,
1450
+ "coverage": 1.0,
1451
+ "false_abstain": 0.0,
1452
+ "null_recall": "n/a",
1453
+ "n": 100
1454
+ },
1455
+ "queue_slow_growth": {
1456
+ "nll": 0.8056600958296468,
1457
+ "brier": 0.46479321367422693,
1458
+ "jsd": 0.2233220382938683,
1459
+ "acc": 0.5,
1460
+ "acc_k": 0.5,
1461
+ "ece": 0.2907683104276657,
1462
+ "sel_acc80": 0.6222222222222222,
1463
+ "coverage": 1.0,
1464
+ "false_abstain": 0.0,
1465
+ "null_recall": "n/a",
1466
+ "n": 56
1467
+ },
1468
+ "cert_30d": {
1469
+ "nll": 0.5765844556293318,
1470
+ "brier": 0.459360806409668,
1471
+ "jsd": 0.1663273119345,
1472
+ "acc": 0.5,
1473
+ "acc_k": 0.5,
1474
+ "ece": 0.34605565667152405,
1475
+ "sel_acc80": 0.6170212765957447,
1476
+ "coverage": 1.0,
1477
+ "false_abstain": 0.0,
1478
+ "null_recall": "n/a",
1479
+ "n": 58
1480
+ },
1481
+ "disk_80pct_48h": {
1482
+ "nll": 0.6364577616833813,
1483
+ "brier": 0.3603717509503686,
1484
+ "jsd": 0.1950357110171284,
1485
+ "acc": 1.0,
1486
+ "acc_k": 1.0,
1487
+ "ece": 0.44600894767791033,
1488
+ "sel_acc80": 1.0,
1489
+ "coverage": 1.0,
1490
+ "false_abstain": 0.0,
1491
+ "null_recall": "n/a",
1492
+ "n": 64
1493
+ },
1494
+ "cpu_high_no_impact": {
1495
+ "nll": 0.7085339514792242,
1496
+ "brier": 0.5101478528535615,
1497
+ "jsd": 0.19822264049400234,
1498
+ "acc": 0.5,
1499
+ "acc_k": 0.5,
1500
+ "ece": 0.37314837455749506,
1501
+ "sel_acc80": 0.625,
1502
+ "coverage": 1.0,
1503
+ "false_abstain": 0.0,
1504
+ "null_recall": "n/a",
1505
+ "n": 50
1506
+ },
1507
+ "single_pod_oom": {
1508
+ "nll": 0.597222631375,
1509
+ "brier": 0.3511328330652838,
1510
+ "jsd": 0.1808235790015383,
1511
+ "acc": 0.5,
1512
+ "acc_k": 0.5,
1513
+ "ece": 0.32174784059707934,
1514
+ "sel_acc80": 0.6190476190476191,
1515
+ "coverage": 1.0,
1516
+ "false_abstain": 0.0,
1517
+ "null_recall": "n/a",
1518
+ "n": 52
1519
+ },
1520
+ "deprecation": {
1521
+ "nll": 1.2131927809820504,
1522
+ "brier": 0.8289228409108279,
1523
+ "jsd": 0.2701922670787307,
1524
+ "acc": 0.5,
1525
+ "acc_k": 0.5,
1526
+ "ece": 0.45578983426094055,
1527
+ "sel_acc80": 0.625,
1528
+ "coverage": 1.0,
1529
+ "false_abstain": 0.0,
1530
+ "null_recall": "n/a",
1531
+ "n": 60
1532
+ },
1533
+ "canary_within_budget": {
1534
+ "nll": 0.68097187510421,
1535
+ "brier": 0.46432714017336996,
1536
+ "jsd": 0.19212515668071925,
1537
+ "acc": 0.5,
1538
+ "acc_k": 0.5,
1539
+ "ece": 0.33959993720054626,
1540
+ "sel_acc80": 0.6153846153846154,
1541
+ "coverage": 1.0,
1542
+ "false_abstain": 0.0,
1543
+ "null_recall": "n/a",
1544
+ "n": 48
1545
+ },
1546
+ "replica_lag_seconds": {
1547
+ "nll": 0.48507738285690005,
1548
+ "brier": 0.2914945314683177,
1549
+ "jsd": 0.14510412619084756,
1550
+ "acc": 0.5,
1551
+ "acc_k": 0.5,
1552
+ "ece": 0.21687236428260803,
1553
+ "sel_acc80": 0.625,
1554
+ "coverage": 1.0,
1555
+ "false_abstain": 0.0,
1556
+ "null_recall": "n/a",
1557
+ "n": 50
1558
+ },
1559
+ "tls_expires_7d": {
1560
+ "nll": 0.8200619861584703,
1561
+ "brier": 0.6263737845037869,
1562
+ "jsd": 0.21875945370810285,
1563
+ "acc": 0.5,
1564
+ "acc_k": 0.5,
1565
+ "ece": 0.41936272382736206,
1566
+ "sel_acc80": 0.6190476190476191,
1567
+ "coverage": 1.0,
1568
+ "false_abstain": 0.0,
1569
+ "null_recall": "n/a",
1570
+ "n": 26
1571
+ },
1572
+ "slow_slo_burn": {
1573
+ "nll": 0.4646857777411605,
1574
+ "brier": 0.2633606501114195,
1575
+ "jsd": 0.15017855189982032,
1576
+ "acc": 1.0,
1577
+ "acc_k": 1.0,
1578
+ "ece": 0.3715018928050995,
1579
+ "sel_acc80": 1.0,
1580
+ "coverage": 1.0,
1581
+ "false_abstain": 0.0,
1582
+ "null_recall": "n/a",
1583
+ "n": 36
1584
+ },
1585
+ "injection_noise": {
1586
+ "nll": 0.3191649156767602,
1587
+ "brier": 0.15667065379041875,
1588
+ "jsd": 0.10271966665831857,
1589
+ "acc": 1.0,
1590
+ "acc_k": 1.0,
1591
+ "ece": 0.2534384096822431,
1592
+ "sel_acc80": 1.0,
1593
+ "coverage": 1.0,
1594
+ "false_abstain": 0.0,
1595
+ "null_recall": "n/a",
1596
+ "n": 310
1597
+ },
1598
+ "health_ok": {
1599
+ "nll": 0.11813930759934398,
1600
+ "brier": 0.039244944056013954,
1601
+ "jsd": 0.03965546739519377,
1602
+ "acc": 1.0,
1603
+ "acc_k": 1.0,
1604
+ "ece": 0.10602798960567297,
1605
+ "sel_acc80": 1.0,
1606
+ "coverage": 1.0,
1607
+ "false_abstain": 0.0,
1608
+ "null_recall": "n/a",
1609
+ "n": 306
1610
+ },
1611
+ "kafka_caught_up": {
1612
+ "nll": 0.1239690909443528,
1613
+ "brier": 0.03284989912051308,
1614
+ "jsd": 0.04179599987809453,
1615
+ "acc": 1.0,
1616
+ "acc_k": 1.0,
1617
+ "ece": 0.11298148923640626,
1618
+ "sel_acc80": 1.0,
1619
+ "coverage": 1.0,
1620
+ "false_abstain": 0.0,
1621
+ "null_recall": "n/a",
1622
+ "n": 278
1623
+ },
1624
+ "invalid_coupon_error": {
1625
+ "nll": 0.4095562707396914,
1626
+ "brier": 0.2070045846973819,
1627
+ "jsd": 0.12855307030768603,
1628
+ "acc": 1.0,
1629
+ "acc_k": 1.0,
1630
+ "ece": 0.3042865756505765,
1631
+ "sel_acc80": 1.0,
1632
+ "coverage": 1.0,
1633
+ "false_abstain": 0.0,
1634
+ "null_recall": "n/a",
1635
+ "n": 322
1636
+ },
1637
+ "stack_trace_handled": {
1638
+ "nll": 0.1369772675048073,
1639
+ "brier": 0.04242081045781862,
1640
+ "jsd": 0.04599428397435028,
1641
+ "acc": 1.0,
1642
+ "acc_k": 1.0,
1643
+ "ece": 0.12313097069062383,
1644
+ "sel_acc80": 1.0,
1645
+ "coverage": 1.0,
1646
+ "false_abstain": 0.0,
1647
+ "null_recall": "n/a",
1648
+ "n": 332
1649
+ },
1650
+ "retry_then_ok": {
1651
+ "nll": 0.23577570062147413,
1652
+ "brier": 0.09354840300021036,
1653
+ "jsd": 0.07807766271472766,
1654
+ "acc": 1.0,
1655
+ "acc_k": 1.0,
1656
+ "ece": 0.20314176128783692,
1657
+ "sel_acc80": 1.0,
1658
+ "coverage": 1.0,
1659
+ "false_abstain": 0.0,
1660
+ "null_recall": "n/a",
1661
+ "n": 284
1662
+ },
1663
+ "feature_flag": {
1664
+ "nll": 0.02957982981648557,
1665
+ "brier": 0.0025294010832958157,
1666
+ "jsd": 0.010185593857561997,
1667
+ "acc": 1.0,
1668
+ "acc_k": 1.0,
1669
+ "ece": 0.028911458685042968,
1670
+ "sel_acc80": 1.0,
1671
+ "coverage": 1.0,
1672
+ "false_abstain": 0.0,
1673
+ "null_recall": "n/a",
1674
+ "n": 282
1675
+ },
1676
+ "cache_hit": {
1677
+ "nll": 0.05577471799801635,
1678
+ "brier": 0.010092673936511587,
1679
+ "jsd": 0.019045345808299963,
1680
+ "acc": 1.0,
1681
+ "acc_k": 1.0,
1682
+ "ece": 0.05296187488349169,
1683
+ "sel_acc80": 1.0,
1684
+ "coverage": 1.0,
1685
+ "false_abstain": 0.0,
1686
+ "null_recall": "n/a",
1687
+ "n": 286
1688
+ },
1689
+ "metrics_scrape": {
1690
+ "nll": 0.19512496383591874,
1691
+ "brier": 0.06495891768903242,
1692
+ "jsd": 0.06532458959189952,
1693
+ "acc": 1.0,
1694
+ "acc_k": 1.0,
1695
+ "ece": 0.17378984264484268,
1696
+ "sel_acc80": 1.0,
1697
+ "coverage": 1.0,
1698
+ "false_abstain": 0.0,
1699
+ "null_recall": "n/a",
1700
+ "n": 328
1701
+ },
1702
+ "cron_ok": {
1703
+ "nll": 0.8260068202421647,
1704
+ "brier": 0.5908376052070468,
1705
+ "jsd": 0.24215122771741,
1706
+ "acc": 0.5,
1707
+ "acc_k": 0.5,
1708
+ "ece": 0.3292480845605173,
1709
+ "sel_acc80": 0.375,
1710
+ "coverage": 1.0,
1711
+ "false_abstain": 0.0,
1712
+ "null_recall": "n/a",
1713
+ "n": 310
1714
+ },
1715
+ "rate_limit_bot": {
1716
+ "nll": 0.2224008145703419,
1717
+ "brier": 0.10465719261617362,
1718
+ "jsd": 0.0723341471248896,
1719
+ "acc": 1.0,
1720
+ "acc_k": 1.0,
1721
+ "ece": 0.18123274389654398,
1722
+ "sel_acc80": 1.0,
1723
+ "coverage": 1.0,
1724
+ "false_abstain": 0.0,
1725
+ "null_recall": "n/a",
1726
+ "n": 256
1727
+ },
1728
+ "http_404": {
1729
+ "nll": 0.0272498202190122,
1730
+ "brier": 0.0025453067235740637,
1731
+ "jsd": 0.009377000999055841,
1732
+ "acc": 1.0,
1733
+ "acc_k": 1.0,
1734
+ "ece": 0.026570967720739525,
1735
+ "sel_acc80": 1.0,
1736
+ "coverage": 1.0,
1737
+ "false_abstain": 0.0,
1738
+ "null_recall": "n/a",
1739
+ "n": 318
1740
+ },
1741
+ "deploy_success": {
1742
+ "nll": 0.11437534991524133,
1743
+ "brier": 0.022894189185427016,
1744
+ "jsd": 0.03898171724942419,
1745
+ "acc": 1.0,
1746
+ "acc_k": 1.0,
1747
+ "ece": 0.10790608627874346,
1748
+ "sel_acc80": 1.0,
1749
+ "coverage": 1.0,
1750
+ "false_abstain": 0.0,
1751
+ "null_recall": "n/a",
1752
+ "n": 244
1753
+ },
1754
+ "user_not_found": {
1755
+ "nll": 0.7467332928805962,
1756
+ "brier": 0.4869266134970554,
1757
+ "jsd": 0.20417162215696066,
1758
+ "acc": 0.5,
1759
+ "acc_k": 0.5,
1760
+ "ece": 0.3224844210812835,
1761
+ "sel_acc80": 0.6228813559322034,
1762
+ "coverage": 1.0,
1763
+ "false_abstain": 0.0,
1764
+ "null_recall": "n/a",
1765
+ "n": 294
1766
+ },
1767
+ "request_ok": {
1768
+ "nll": 0.15962773605536434,
1769
+ "brier": 0.045111188981216126,
1770
+ "jsd": 0.05396733304793474,
1771
+ "acc": 1.0,
1772
+ "acc_k": 1.0,
1773
+ "ece": 0.14663683669907707,
1774
+ "sel_acc80": 1.0,
1775
+ "coverage": 1.0,
1776
+ "false_abstain": 0.0,
1777
+ "null_recall": "n/a",
1778
+ "n": 350
1779
+ }
1780
+ },
1781
+ "by_qtype": {
1782
+ "choice": {
1783
+ "nll": 0.48022365288831015,
1784
+ "brier": 0.25716945445373307,
1785
+ "jsd": 0.14152253563996733,
1786
+ "acc": 0.851,
1787
+ "acc_k": 0.851,
1788
+ "ece": 0.13448732795317966,
1789
+ "sel_acc80": 0.905,
1790
+ "coverage": 1.0,
1791
+ "false_abstain": 0.0,
1792
+ "null_recall": "n/a",
1793
+ "n": 3000
1794
+ },
1795
+ "noul": {
1796
+ "nll": 0.31319126489478805,
1797
+ "brier": 0.20287533336799038,
1798
+ "jsd": 0.0849231311763303,
1799
+ "acc": 0.825,
1800
+ "acc_k": 0.825,
1801
+ "ece": 0.10489587610960006,
1802
+ "sel_acc80": 0.9566666666666667,
1803
+ "coverage": 1.0,
1804
+ "false_abstain": 0.0,
1805
+ "null_recall": "n/a",
1806
+ "n": 3000
1807
+ }
1808
+ },
1809
+ "by_label_class": {
1810
+ "answer": {
1811
+ "nll": 0.396707458891549,
1812
+ "brier": 0.23002239391086174,
1813
+ "jsd": 0.11322283340814882,
1814
+ "acc": 0.838,
1815
+ "acc_k": 0.838,
1816
+ "ece": 0.07345882829030351,
1817
+ "sel_acc80": 0.8979166666666667,
1818
+ "coverage": 1.0,
1819
+ "false_abstain": 0.0,
1820
+ "null_recall": "n/a",
1821
+ "n": 6000
1822
+ }
1823
+ },
1824
+ "baselines": {
1825
+ "majority_class": 0,
1826
+ "majority_acc": 0.7916666666666666
1827
+ }
1828
+ },
1829
+ "systemone_lite_hard": {
1830
+ "n": 5400,
1831
+ "n_total": 5400,
1832
+ "suffix_overflow": 0,
1833
+ "raw": {
1834
+ "nll": 0.43450476654815856,
1835
+ "brier": 0.23583658134544866,
1836
+ "jsd": 0.11106552302444929,
1837
+ "acc": 0.850925925925926,
1838
+ "acc_k": 0.850925925925926,
1839
+ "ece": 0.05852483691440687,
1840
+ "sel_acc80": 0.8976851851851851,
1841
+ "coverage": 1.0,
1842
+ "false_abstain": 0.0,
1843
+ "null_recall": "n/a"
1844
+ },
1845
+ "by_family": {
1846
+ "systemone_lite": {
1847
+ "nll": 0.43450476654815856,
1848
+ "brier": 0.23583658134544866,
1849
+ "jsd": 0.11106552302444929,
1850
+ "acc": 0.850925925925926,
1851
+ "acc_k": 0.850925925925926,
1852
+ "ece": 0.05852483691440687,
1853
+ "sel_acc80": 0.8976851851851851,
1854
+ "coverage": 1.0,
1855
+ "false_abstain": 0.0,
1856
+ "null_recall": "n/a",
1857
+ "n": 5400
1858
+ }
1859
+ },
1860
+ "by_qtype": {
1861
+ "choice": {
1862
+ "nll": 0.47634740362930744,
1863
+ "brier": 0.22078883830669857,
1864
+ "jsd": 0.0890289134343277,
1865
+ "acc": 0.8522222222222222,
1866
+ "acc_k": 0.8522222222222222,
1867
+ "ece": 0.06805240516861279,
1868
+ "sel_acc80": 0.9319444444444445,
1869
+ "coverage": 1.0,
1870
+ "false_abstain": 0.0,
1871
+ "null_recall": "n/a",
1872
+ "n": 1800
1873
+ },
1874
+ "noul": {
1875
+ "nll": 0.23199631304256532,
1876
+ "brier": 0.14232818861742158,
1877
+ "jsd": 0.06226400940582674,
1878
+ "acc": 0.8994444444444445,
1879
+ "acc_k": 0.8994444444444445,
1880
+ "ece": 0.01822677158647116,
1881
+ "sel_acc80": 0.9597222222222223,
1882
+ "coverage": 1.0,
1883
+ "false_abstain": 0.0,
1884
+ "null_recall": "n/a",
1885
+ "n": 1800
1886
+ },
1887
+ "score": {
1888
+ "nll": 0.595170582972603,
1889
+ "brier": 0.3443927171122257,
1890
+ "jsd": 0.18190364623319336,
1891
+ "acc": 0.8011111111111111,
1892
+ "acc_k": 0.8011111111111111,
1893
+ "ece": 0.17747255620029237,
1894
+ "sel_acc80": 0.8513888888888889,
1895
+ "coverage": 1.0,
1896
+ "false_abstain": 0.0,
1897
+ "null_recall": "n/a",
1898
+ "n": 1800
1899
+ }
1900
+ },
1901
+ "by_label_class": {
1902
+ "answer": {
1903
+ "nll": 0.43450476654815856,
1904
+ "brier": 0.23583658134544866,
1905
+ "jsd": 0.11106552302444929,
1906
+ "acc": 0.850925925925926,
1907
+ "acc_k": 0.850925925925926,
1908
+ "ece": 0.05852483691440687,
1909
+ "sel_acc80": 0.8976851851851851,
1910
+ "coverage": 1.0,
1911
+ "false_abstain": 0.0,
1912
+ "null_recall": "n/a",
1913
+ "n": 5400
1914
+ }
1915
+ },
1916
+ "baselines": {
1917
+ "majority_class": 1,
1918
+ "majority_acc": 0.44166666666666665
1919
+ }
1920
+ },
1921
+ "tree_choice_cap": {
1922
+ "n": 720,
1923
+ "n_total": 720,
1924
+ "suffix_overflow": 0,
1925
+ "raw": {
1926
+ "nll": 1.0937767414927673,
1927
+ "brier": 0.4118990134350543,
1928
+ "jsd": 0.2282905942697906,
1929
+ "acc": 0.6902777777777778,
1930
+ "acc_k": 0.7319444444444444,
1931
+ "ece": 0.07905487779838341,
1932
+ "sel_acc80": 0.7934027777777778,
1933
+ "coverage": 0.8611111111111112,
1934
+ "false_abstain": 0.1388888888888889,
1935
+ "null_recall": "n/a"
1936
+ },
1937
+ "by_family": {
1938
+ "incident_catalog": {
1939
+ "nll": 1.0937767414927673,
1940
+ "brier": 0.4118990134350543,
1941
+ "jsd": 0.2282905942697906,
1942
+ "acc": 0.6902777777777778,
1943
+ "acc_k": 0.7319444444444444,
1944
+ "ece": 0.07905487779838341,
1945
+ "sel_acc80": 0.7934027777777778,
1946
+ "coverage": 0.8611111111111112,
1947
+ "false_abstain": 0.1388888888888889,
1948
+ "null_recall": "n/a",
1949
+ "n": 720
1950
+ }
1951
+ },
1952
+ "by_qtype": {
1953
+ "choice": {
1954
+ "nll": 1.0937767414927673,
1955
+ "brier": 0.4118990134350543,
1956
+ "jsd": 0.2282905942697906,
1957
+ "acc": 0.6902777777777778,
1958
+ "acc_k": 0.7319444444444444,
1959
+ "ece": 0.07905487779838341,
1960
+ "sel_acc80": 0.7934027777777778,
1961
+ "coverage": 0.8611111111111112,
1962
+ "false_abstain": 0.1388888888888889,
1963
+ "null_recall": "n/a",
1964
+ "n": 720
1965
+ }
1966
+ },
1967
+ "by_label_class": {
1968
+ "answer": {
1969
+ "nll": 1.0937767414927673,
1970
+ "brier": 0.4118990134350543,
1971
+ "jsd": 0.2282905942697906,
1972
+ "acc": 0.6902777777777778,
1973
+ "acc_k": 0.7319444444444444,
1974
+ "ece": 0.07905487779838341,
1975
+ "sel_acc80": 0.7934027777777778,
1976
+ "coverage": 0.8611111111111112,
1977
+ "false_abstain": 0.1388888888888889,
1978
+ "null_recall": "n/a",
1979
+ "n": 720
1980
+ }
1981
+ },
1982
+ "baselines": {
1983
+ "majority_class": 3,
1984
+ "majority_acc": 0.11388888888888889
1985
+ }
1986
+ },
1987
+ "typed_decisions_test": {
1988
+ "n": 2000,
1989
+ "n_total": 2000,
1990
+ "suffix_overflow": 0,
1991
+ "raw": {
1992
+ "nll": 1.183971535602263,
1993
+ "brier": 0.21124234378656148,
1994
+ "jsd": 0.08978436233492346,
1995
+ "acc": 0.5325,
1996
+ "acc_k": 0.5325,
1997
+ "ece": 0.0960818936675787,
1998
+ "sel_acc80": 0.55125,
1999
+ "coverage": 1.0,
2000
+ "false_abstain": 0.0,
2001
+ "null_recall": "n/a"
2002
+ },
2003
+ "by_family": {
2004
+ "agent_trace_observability": {
2005
+ "nll": 1.4513733308753711,
2006
+ "brier": 0.27746615896441984,
2007
+ "jsd": 0.11437914671823042,
2008
+ "acc": 0.4,
2009
+ "acc_k": 0.4,
2010
+ "ece": 0.21149230748414996,
2011
+ "sel_acc80": 0.4125,
2012
+ "coverage": 1.0,
2013
+ "false_abstain": 0.0,
2014
+ "null_recall": "n/a",
2015
+ "n": 500
2016
+ },
2017
+ "customer_service": {
2018
+ "nll": 1.1394922824048543,
2019
+ "brier": 0.1685314240724029,
2020
+ "jsd": 0.07635420781877246,
2021
+ "acc": 0.638,
2022
+ "acc_k": 0.638,
2023
+ "ece": 0.09256491619348527,
2024
+ "sel_acc80": 0.6825,
2025
+ "coverage": 1.0,
2026
+ "false_abstain": 0.0,
2027
+ "null_recall": "n/a",
2028
+ "n": 500
2029
+ },
2030
+ "invoice_processing": {
2031
+ "nll": 1.0241192434455133,
2032
+ "brier": 0.25477993975703184,
2033
+ "jsd": 0.11173192033283712,
2034
+ "acc": 0.582,
2035
+ "acc_k": 0.582,
2036
+ "ece": 0.08850071275234224,
2037
+ "sel_acc80": 0.64,
2038
+ "coverage": 1.0,
2039
+ "false_abstain": 0.0,
2040
+ "null_recall": "n/a",
2041
+ "n": 500
2042
+ },
2043
+ "security_incidents": {
2044
+ "nll": 1.1209012856833143,
2045
+ "brier": 0.14419185235239126,
2046
+ "jsd": 0.056672174469853814,
2047
+ "acc": 0.51,
2048
+ "acc_k": 0.51,
2049
+ "ece": 0.1298270362019539,
2050
+ "sel_acc80": 0.505,
2051
+ "coverage": 1.0,
2052
+ "false_abstain": 0.0,
2053
+ "null_recall": "n/a",
2054
+ "n": 500
2055
+ }
2056
+ },
2057
+ "by_qtype": {
2058
+ "choice": {
2059
+ "nll": 1.4624093197198917,
2060
+ "brier": 0.22882638324834792,
2061
+ "jsd": 0.10929266556165232,
2062
+ "acc": 0.49166666666666664,
2063
+ "acc_k": 0.49166666666666664,
2064
+ "ece": 0.12719455897808077,
2065
+ "sel_acc80": 0.5333333333333333,
2066
+ "coverage": 1.0,
2067
+ "false_abstain": 0.0,
2068
+ "null_recall": "n/a",
2069
+ "n": 600
2070
+ },
2071
+ "noul": {
2072
+ "nll": 0.8045478477317457,
2073
+ "brier": 0.25816601068728884,
2074
+ "jsd": 0.08310478332934974,
2075
+ "acc": 0.5783333333333334,
2076
+ "acc_k": 0.5783333333333334,
2077
+ "ece": 0.15133020063241323,
2078
+ "sel_acc80": 0.5895833333333333,
2079
+ "coverage": 1.0,
2080
+ "false_abstain": 0.0,
2081
+ "null_recall": "n/a",
2082
+ "n": 600
2083
+ },
2084
+ "score": {
2085
+ "nll": 1.2597109634169303,
2086
+ "brier": 0.16286156401467608,
2087
+ "jsd": 0.08016281916905708,
2088
+ "acc": 0.52875,
2089
+ "acc_k": 0.52875,
2090
+ "ece": 0.07651440937072038,
2091
+ "sel_acc80": 0.55,
2092
+ "coverage": 1.0,
2093
+ "false_abstain": 0.0,
2094
+ "null_recall": "n/a",
2095
+ "n": 800
2096
+ }
2097
+ },
2098
+ "by_label_class": {
2099
+ "answer": {
2100
+ "nll": 1.183971535602263,
2101
+ "brier": 0.21124234378656148,
2102
+ "jsd": 0.08978436233492346,
2103
+ "acc": 0.5325,
2104
+ "acc_k": 0.5325,
2105
+ "ece": 0.0960818936675787,
2106
+ "sel_acc80": 0.55125,
2107
+ "coverage": 1.0,
2108
+ "false_abstain": 0.0,
2109
+ "null_recall": "n/a",
2110
+ "n": 2000
2111
+ }
2112
+ },
2113
+ "baselines": {
2114
+ "majority_class": 1,
2115
+ "majority_acc": 0.2905
2116
+ }
2117
+ },
2118
+ "typed_decisions_train": {
2119
+ "n": 6000,
2120
+ "n_total": 6000,
2121
+ "suffix_overflow": 0,
2122
+ "raw": {
2123
+ "nll": 1.1985372981046378,
2124
+ "brier": 0.2329991959274657,
2125
+ "jsd": 0.09837575798408288,
2126
+ "acc": 0.5223333333333333,
2127
+ "acc_k": 0.5223333333333333,
2128
+ "ece": 0.10562264278034368,
2129
+ "sel_acc80": 0.5375,
2130
+ "coverage": 1.0,
2131
+ "false_abstain": 0.0,
2132
+ "null_recall": "n/a"
2133
+ },
2134
+ "by_family": {
2135
+ "agent_trace_observability": {
2136
+ "nll": 1.416481477970894,
2137
+ "brier": 0.25422688167917284,
2138
+ "jsd": 0.10777658676024122,
2139
+ "acc": 0.4266666666666667,
2140
+ "acc_k": 0.4266666666666667,
2141
+ "ece": 0.18357536943753558,
2142
+ "sel_acc80": 0.42916666666666664,
2143
+ "coverage": 1.0,
2144
+ "false_abstain": 0.0,
2145
+ "null_recall": "n/a",
2146
+ "n": 1500
2147
+ },
2148
+ "customer_service": {
2149
+ "nll": 1.1214988914219903,
2150
+ "brier": 0.19128662847914574,
2151
+ "jsd": 0.08370220742725006,
2152
+ "acc": 0.6106666666666667,
2153
+ "acc_k": 0.6106666666666667,
2154
+ "ece": 0.08198627370595929,
2155
+ "sel_acc80": 0.6433333333333333,
2156
+ "coverage": 1.0,
2157
+ "false_abstain": 0.0,
2158
+ "null_recall": "n/a",
2159
+ "n": 1500
2160
+ },
2161
+ "invoice_processing": {
2162
+ "nll": 1.1031419902595583,
2163
+ "brier": 0.3050021381875509,
2164
+ "jsd": 0.13027031117221108,
2165
+ "acc": 0.5333333333333333,
2166
+ "acc_k": 0.5333333333333333,
2167
+ "ece": 0.10525429819027585,
2168
+ "sel_acc80": 0.575,
2169
+ "coverage": 1.0,
2170
+ "false_abstain": 0.0,
2171
+ "null_recall": "n/a",
2172
+ "n": 1500
2173
+ },
2174
+ "security_incidents": {
2175
+ "nll": 1.1530268327661088,
2176
+ "brier": 0.1814811353639933,
2177
+ "jsd": 0.07175392657662916,
2178
+ "acc": 0.5186666666666667,
2179
+ "acc_k": 0.5186666666666667,
2180
+ "ece": 0.12120020204782486,
2181
+ "sel_acc80": 0.5258333333333334,
2182
+ "coverage": 1.0,
2183
+ "false_abstain": 0.0,
2184
+ "null_recall": "n/a",
2185
+ "n": 1500
2186
+ }
2187
+ },
2188
+ "by_qtype": {
2189
+ "choice": {
2190
+ "nll": 1.4635702511121844,
2191
+ "brier": 0.2532055828202569,
2192
+ "jsd": 0.11942735571638688,
2193
+ "acc": 0.4961111111111111,
2194
+ "acc_k": 0.4961111111111111,
2195
+ "ece": 0.12604168651832473,
2196
+ "sel_acc80": 0.5368055555555555,
2197
+ "coverage": 1.0,
2198
+ "false_abstain": 0.0,
2199
+ "null_recall": "n/a",
2200
+ "n": 1800
2201
+ },
2202
+ "noul": {
2203
+ "nll": 0.8394133296564307,
2204
+ "brier": 0.2816899204625882,
2205
+ "jsd": 0.09059050796156125,
2206
+ "acc": 0.5633333333333334,
2207
+ "acc_k": 0.5633333333333334,
2208
+ "ece": 0.1775682842400339,
2209
+ "sel_acc80": 0.5701388888888889,
2210
+ "coverage": 1.0,
2211
+ "false_abstain": 0.0,
2212
+ "null_recall": "n/a",
2213
+ "n": 1800
2214
+ },
2215
+ "score": {
2216
+ "nll": 1.2691055596851333,
2217
+ "brier": 0.18132636235653043,
2218
+ "jsd": 0.08842599720174607,
2219
+ "acc": 0.51125,
2220
+ "acc_k": 0.51125,
2221
+ "ece": 0.08015046964089076,
2222
+ "sel_acc80": 0.5317708333333333,
2223
+ "coverage": 1.0,
2224
+ "false_abstain": 0.0,
2225
+ "null_recall": "n/a",
2226
+ "n": 2400
2227
+ }
2228
+ },
2229
+ "by_label_class": {
2230
+ "answer": {
2231
+ "nll": 1.1985372981046378,
2232
+ "brier": 0.2329991959274657,
2233
+ "jsd": 0.09837575798408288,
2234
+ "acc": 0.5223333333333333,
2235
+ "acc_k": 0.5223333333333333,
2236
+ "ece": 0.10562264278034368,
2237
+ "sel_acc80": 0.5375,
2238
+ "coverage": 1.0,
2239
+ "false_abstain": 0.0,
2240
+ "null_recall": "n/a",
2241
+ "n": 6000
2242
+ }
2243
+ },
2244
+ "baselines": {
2245
+ "majority_class": 1,
2246
+ "majority_acc": 0.3105
2247
+ }
2248
+ },
2249
+ "wh_heldout_family": {
2250
+ "n": 422,
2251
+ "n_total": 422,
2252
+ "suffix_overflow": 0,
2253
+ "raw": {
2254
+ "nll": 0.3240292072876919,
2255
+ "brier": 0.17772530751240193,
2256
+ "jsd": 0.07779958816769972,
2257
+ "acc": 0.8744075829383886,
2258
+ "acc_k": 0.9045092838196287,
2259
+ "ece": 0.04873869755256795,
2260
+ "sel_acc80": 0.9437869822485208,
2261
+ "conf_wrong": 0.0,
2262
+ "coverage": 0.8625592417061612,
2263
+ "false_abstain": 0.04774535809018567,
2264
+ "null_recall": 0.8888888888888888,
2265
+ "auroc_null": 0.9826112584733274,
2266
+ "auroc_conf": 0.9737105806071323
2267
+ },
2268
+ "by_family": {
2269
+ "procurement_request": {
2270
+ "nll": 0.31726594171575667,
2271
+ "brier": 0.15641665156788495,
2272
+ "jsd": 0.06340224526164076,
2273
+ "acc": 0.9014084507042254,
2274
+ "acc_k": 0.9375,
2275
+ "ece": 0.07553887535148941,
2276
+ "sel_acc80": 0.956140350877193,
2277
+ "conf_wrong": 0.0,
2278
+ "coverage": 0.8802816901408451,
2279
+ "false_abstain": 0.0390625,
2280
+ "null_recall": 0.8571428571428571,
2281
+ "auroc_null": 0.9827008928571429,
2282
+ "auroc_conf": 0.9782366071428572,
2283
+ "n": 142
2284
+ },
2285
+ "research_ethics_review": {
2286
+ "nll": 0.41234907161332957,
2287
+ "brier": 0.23254687981548977,
2288
+ "jsd": 0.10529413759841766,
2289
+ "acc": 0.8396946564885496,
2290
+ "acc_k": 0.8376068376068376,
2291
+ "ece": 0.07992691961863568,
2292
+ "sel_acc80": 0.9238095238095239,
2293
+ "conf_wrong": 0.0,
2294
+ "coverage": 0.8549618320610687,
2295
+ "false_abstain": 0.042735042735042736,
2296
+ "null_recall": 1.0,
2297
+ "auroc_null": 0.9865689865689865,
2298
+ "auroc_conf": 0.9847374847374847,
2299
+ "n": 131
2300
+ },
2301
+ "subscription_cancellation": {
2302
+ "nll": 0.25282438503639165,
2303
+ "brier": 0.14983405363600533,
2304
+ "jsd": 0.06734748559881597,
2305
+ "acc": 0.8791946308724832,
2306
+ "acc_k": 0.9318181818181818,
2307
+ "ece": 0.04965120893996835,
2308
+ "sel_acc80": 0.9666666666666667,
2309
+ "conf_wrong": 0.0,
2310
+ "coverage": 0.8523489932885906,
2311
+ "false_abstain": 0.06060606060606061,
2312
+ "null_recall": 0.8235294117647058,
2313
+ "auroc_null": 0.9795008912655971,
2314
+ "auroc_conf": 0.9714795008912657,
2315
+ "n": 149
2316
+ }
2317
+ },
2318
+ "by_qtype": {},
2319
+ "by_label_class": {
2320
+ "answer": {
2321
+ "nll": 0.3270902509793986,
2322
+ "brier": 0.1794731921212136,
2323
+ "jsd": 0.07638232337116445,
2324
+ "acc": 0.8726790450928382,
2325
+ "acc_k": 0.9045092838196287,
2326
+ "ece": 0.04732005681536247,
2327
+ "sel_acc80": 0.9403973509933775,
2328
+ "coverage": 0.9522546419098143,
2329
+ "false_abstain": 0.04774535809018567,
2330
+ "null_recall": "n/a",
2331
+ "n": 377
2332
+ },
2333
+ "null": {
2334
+ "nll": 0.2983844634705046,
2335
+ "brier": 0.16308191867857966,
2336
+ "jsd": 0.08967311768533967,
2337
+ "acc": 0.8888888888888888,
2338
+ "ece": 0.13434556325276692,
2339
+ "sel_acc80": 0.9722222222222222,
2340
+ "conf_wrong": 0.0,
2341
+ "coverage": 0.11111111111111116,
2342
+ "false_abstain": "n/a",
2343
+ "null_recall": 0.8888888888888888,
2344
+ "n": 45
2345
+ }
2346
+ },
2347
+ "baselines": {
2348
+ "majority_class": 0,
2349
+ "majority_acc": 0.38388625592417064
2350
+ }
2351
+ },
2352
+ "wh_heldout_grammar": {
2353
+ "n": 816,
2354
+ "n_total": 816,
2355
+ "suffix_overflow": 0,
2356
+ "raw": {
2357
+ "nll": 0.2472013947656661,
2358
+ "brier": 0.14714674629576405,
2359
+ "jsd": 0.0655881649228148,
2360
+ "acc": 0.8933823529411765,
2361
+ "acc_k": 0.925414364640884,
2362
+ "ece": 0.02050217598968861,
2363
+ "sel_acc80": 0.9647779479326187,
2364
+ "conf_wrong": 0.0,
2365
+ "coverage": 0.8455882352941176,
2366
+ "false_abstain": 0.05386740331491713,
2367
+ "null_recall": 0.9456521739130435,
2368
+ "auroc_null": 0.9844012731203459,
2369
+ "auroc_conf": 0.9806629834254144
2370
+ },
2371
+ "by_family": {
2372
+ "visa_extension": {
2373
+ "nll": 0.3876584563899138,
2374
+ "brier": 0.23389678729229724,
2375
+ "jsd": 0.10288573687884103,
2376
+ "acc": 0.851063829787234,
2377
+ "acc_k": 0.8782608695652174,
2378
+ "ece": 0.0661268077843578,
2379
+ "sel_acc80": 0.9203539823008849,
2380
+ "conf_wrong": 0.0,
2381
+ "coverage": 0.7588652482269503,
2382
+ "false_abstain": 0.0782608695652174,
2383
+ "null_recall": 0.9615384615384616,
2384
+ "auroc_null": 0.9709030100334448,
2385
+ "auroc_conf": 0.9692307692307692,
2386
+ "n": 141
2387
+ },
2388
+ "loan_approval": {
2389
+ "nll": 7.454925248028152e-05,
2390
+ "brier": 1.315695971451629e-07,
2391
+ "jsd": 2.5833865575709257e-05,
2392
+ "acc": 1.0,
2393
+ "acc_k": 1.0,
2394
+ "ece": 7.45158125884382e-05,
2395
+ "sel_acc80": 1.0,
2396
+ "conf_wrong": 0.0,
2397
+ "coverage": 0.9708029197080292,
2398
+ "false_abstain": 0.0,
2399
+ "null_recall": 1.0,
2400
+ "auroc_null": 1.0,
2401
+ "auroc_conf": 1.0,
2402
+ "n": 137
2403
+ },
2404
+ "content_moderation": {
2405
+ "nll": 0.5935040969149519,
2406
+ "brier": 0.3596948371934755,
2407
+ "jsd": 0.16132759921807785,
2408
+ "acc": 0.703125,
2409
+ "acc_k": 0.8247422680412371,
2410
+ "ece": 0.10629645804874599,
2411
+ "sel_acc80": 0.7864077669902912,
2412
+ "conf_wrong": 0.0,
2413
+ "coverage": 0.5703125,
2414
+ "false_abstain": 0.26804123711340205,
2415
+ "null_recall": 0.9354838709677419,
2416
+ "auroc_null": 0.9002327901563019,
2417
+ "auroc_conf": 0.9068839374792153,
2418
+ "n": 128
2419
+ },
2420
+ "warranty_claim": {
2421
+ "nll": 0.04521754519140885,
2422
+ "brier": 0.030164132379325935,
2423
+ "jsd": 0.012534975904067302,
2424
+ "acc": 0.971830985915493,
2425
+ "acc_k": 0.9710144927536232,
2426
+ "ece": 0.025068008563887925,
2427
+ "sel_acc80": 1.0,
2428
+ "conf_wrong": 0.0,
2429
+ "coverage": 0.971830985915493,
2430
+ "false_abstain": 0.0,
2431
+ "null_recall": 1.0,
2432
+ "auroc_null": 1.0,
2433
+ "auroc_conf": 1.0,
2434
+ "n": 142
2435
+ },
2436
+ "shift_swap": {
2437
+ "nll": 0.31995094781288197,
2438
+ "brier": 0.1809786294088722,
2439
+ "jsd": 0.07359853995759635,
2440
+ "acc": 0.8615384615384616,
2441
+ "acc_k": 0.8907563025210085,
2442
+ "ece": 0.054814853805762016,
2443
+ "sel_acc80": 0.9423076923076923,
2444
+ "conf_wrong": 0.0,
2445
+ "coverage": 0.9,
2446
+ "false_abstain": 0.03361344537815126,
2447
+ "null_recall": 0.8181818181818182,
2448
+ "auroc_null": 0.985485103132162,
2449
+ "auroc_conf": 0.9572192513368983,
2450
+ "n": 130
2451
+ },
2452
+ "vendor_onboarding": {
2453
+ "nll": 0.16712495284237217,
2454
+ "brier": 0.09594791421965362,
2455
+ "jsd": 0.05081018090389403,
2456
+ "acc": 0.9565217391304348,
2457
+ "acc_k": 0.9508196721311475,
2458
+ "ece": 0.07468755521636082,
2459
+ "sel_acc80": 0.990990990990991,
2460
+ "conf_wrong": 0.0,
2461
+ "coverage": 0.8840579710144928,
2462
+ "false_abstain": 0.0,
2463
+ "null_recall": 1.0,
2464
+ "auroc_null": 1.0,
2465
+ "auroc_conf": 1.0,
2466
+ "n": 138
2467
+ }
2468
+ },
2469
+ "by_qtype": {},
2470
+ "by_label_class": {
2471
+ "null": {
2472
+ "nll": 0.277453074058921,
2473
+ "brier": 0.14561172018866156,
2474
+ "jsd": 0.08687484324735249,
2475
+ "acc": 0.9456521739130435,
2476
+ "ece": 0.1358135566115379,
2477
+ "sel_acc80": 0.972972972972973,
2478
+ "conf_wrong": 0.0,
2479
+ "coverage": 0.05434782608695654,
2480
+ "false_abstain": "n/a",
2481
+ "null_recall": 0.9456521739130435,
2482
+ "n": 92
2483
+ },
2484
+ "answer": {
2485
+ "nll": 0.24335725872287678,
2486
+ "brier": 0.14734180486185994,
2487
+ "jsd": 0.06288322789814979,
2488
+ "acc": 0.8867403314917127,
2489
+ "acc_k": 0.925414364640884,
2490
+ "ece": 0.03288542010013568,
2491
+ "sel_acc80": 0.9706896551724138,
2492
+ "coverage": 0.9461325966850829,
2493
+ "false_abstain": 0.05386740331491713,
2494
+ "null_recall": "n/a",
2495
+ "n": 724
2496
+ }
2497
+ },
2498
+ "baselines": {
2499
+ "majority_class": 0,
2500
+ "majority_acc": 0.38480392156862747
2501
+ }
2502
+ },
2503
+ "wh_heldout_style": {
2504
+ "n": 403,
2505
+ "n_total": 403,
2506
+ "suffix_overflow": 0,
2507
+ "raw": {
2508
+ "nll": 0.26691198757004164,
2509
+ "brier": 0.1464047042686628,
2510
+ "jsd": 0.0662673268053048,
2511
+ "acc": 0.890818858560794,
2512
+ "acc_k": 0.890818858560794,
2513
+ "ece": 0.031059944674039948,
2514
+ "sel_acc80": 0.9845201238390093,
2515
+ "coverage": 1.0,
2516
+ "false_abstain": 0.0,
2517
+ "null_recall": "n/a"
2518
+ },
2519
+ "by_family": {
2520
+ "feature_flag_rollout": {
2521
+ "nll": 0.19010573836138,
2522
+ "brier": 0.11432481201229953,
2523
+ "jsd": 0.05277888022237791,
2524
+ "acc": 0.9090909090909091,
2525
+ "acc_k": 0.9090909090909091,
2526
+ "ece": 0.039632322571494384,
2527
+ "sel_acc80": 1.0,
2528
+ "coverage": 1.0,
2529
+ "false_abstain": 0.0,
2530
+ "null_recall": "n/a",
2531
+ "n": 55
2532
+ },
2533
+ "shift_swap": {
2534
+ "nll": 0.3043634096913596,
2535
+ "brier": 0.1503282508374881,
2536
+ "jsd": 0.06654764978820293,
2537
+ "acc": 0.8939393939393939,
2538
+ "acc_k": 0.8939393939393939,
2539
+ "ece": 0.062232637044155215,
2540
+ "sel_acc80": 0.9811320754716981,
2541
+ "coverage": 1.0,
2542
+ "false_abstain": 0.0,
2543
+ "null_recall": "n/a",
2544
+ "n": 66
2545
+ },
2546
+ "content_moderation": {
2547
+ "nll": 0.4789173250257905,
2548
+ "brier": 0.26149167112429633,
2549
+ "jsd": 0.1279300537119714,
2550
+ "acc": 0.813953488372093,
2551
+ "acc_k": 0.813953488372093,
2552
+ "ece": 0.06183511365291682,
2553
+ "sel_acc80": 0.8571428571428571,
2554
+ "coverage": 1.0,
2555
+ "false_abstain": 0.0,
2556
+ "null_recall": "n/a",
2557
+ "n": 43
2558
+ },
2559
+ "insurance_claim": {
2560
+ "nll": 0.29568089754215104,
2561
+ "brier": 0.16360065632472767,
2562
+ "jsd": 0.06646606171635495,
2563
+ "acc": 0.8529411764705882,
2564
+ "acc_k": 0.8529411764705882,
2565
+ "ece": 0.1020367794177111,
2566
+ "sel_acc80": 1.0,
2567
+ "coverage": 1.0,
2568
+ "false_abstain": 0.0,
2569
+ "null_recall": "n/a",
2570
+ "n": 34
2571
+ },
2572
+ "visa_extension": {
2573
+ "nll": 0.09558577342528488,
2574
+ "brier": 0.04933176843684583,
2575
+ "jsd": 0.02922906947317498,
2576
+ "acc": 0.972972972972973,
2577
+ "acc_k": 0.972972972972973,
2578
+ "ece": 0.03434973310779885,
2579
+ "sel_acc80": 1.0,
2580
+ "coverage": 1.0,
2581
+ "false_abstain": 0.0,
2582
+ "null_recall": "n/a",
2583
+ "n": 37
2584
+ },
2585
+ "warranty_claim": {
2586
+ "nll": 0.31847504141119937,
2587
+ "brier": 0.1722281656536195,
2588
+ "jsd": 0.07974897957552128,
2589
+ "acc": 0.8813559322033898,
2590
+ "acc_k": 0.8813559322033898,
2591
+ "ece": 0.03718661453764316,
2592
+ "sel_acc80": 0.9791666666666666,
2593
+ "coverage": 1.0,
2594
+ "false_abstain": 0.0,
2595
+ "null_recall": "n/a",
2596
+ "n": 59
2597
+ },
2598
+ "loan_approval": {
2599
+ "nll": 0.12007633533318733,
2600
+ "brier": 0.06839648073371568,
2601
+ "jsd": 0.027442892286818,
2602
+ "acc": 0.9583333333333334,
2603
+ "acc_k": 0.9583333333333334,
2604
+ "ece": 0.04480526472131414,
2605
+ "sel_acc80": 1.0,
2606
+ "coverage": 1.0,
2607
+ "false_abstain": 0.0,
2608
+ "null_recall": "n/a",
2609
+ "n": 48
2610
+ },
2611
+ "vendor_onboarding": {
2612
+ "nll": 0.299750322830979,
2613
+ "brier": 0.17565957622934011,
2614
+ "jsd": 0.07452437378826761,
2615
+ "acc": 0.8524590163934426,
2616
+ "acc_k": 0.8524590163934426,
2617
+ "ece": 0.08769017262536971,
2618
+ "sel_acc80": 0.9591836734693877,
2619
+ "coverage": 1.0,
2620
+ "false_abstain": 0.0,
2621
+ "null_recall": "n/a",
2622
+ "n": 61
2623
+ }
2624
+ },
2625
+ "by_qtype": {},
2626
+ "by_label_class": {
2627
+ "answer": {
2628
+ "nll": 0.26691198757004164,
2629
+ "brier": 0.1464047042686628,
2630
+ "jsd": 0.0662673268053048,
2631
+ "acc": 0.890818858560794,
2632
+ "acc_k": 0.890818858560794,
2633
+ "ece": 0.031059944674039948,
2634
+ "sel_acc80": 0.9845201238390093,
2635
+ "coverage": 1.0,
2636
+ "false_abstain": 0.0,
2637
+ "null_recall": "n/a",
2638
+ "n": 403
2639
+ }
2640
+ },
2641
+ "baselines": {
2642
+ "majority_class": 1,
2643
+ "majority_acc": 0.4267990074441687
2644
+ }
2645
+ },
2646
+ "wh_level7": {
2647
+ "n": 880,
2648
+ "n_total": 880,
2649
+ "suffix_overflow": 0,
2650
+ "raw": {
2651
+ "nll": 0.8425315491747045,
2652
+ "brier": 0.566066637960518,
2653
+ "jsd": 0.22709053783966102,
2654
+ "acc": 0.5443181818181818,
2655
+ "acc_k": 0.5443181818181818,
2656
+ "ece": 0.17712931768460705,
2657
+ "sel_acc80": 0.5738636363636364,
2658
+ "coverage": 1.0,
2659
+ "false_abstain": 0.0,
2660
+ "null_recall": "n/a"
2661
+ },
2662
+ "by_family": {
2663
+ "temporal": {
2664
+ "nll": 0.7525894214290174,
2665
+ "brier": 0.5283106465099253,
2666
+ "jsd": 0.20930622376758312,
2667
+ "acc": 0.5727272727272728,
2668
+ "acc_k": 0.5727272727272728,
2669
+ "ece": 0.18158261044458907,
2670
+ "sel_acc80": 0.6193181818181818,
2671
+ "coverage": 1.0,
2672
+ "false_abstain": 0.0,
2673
+ "null_recall": "n/a",
2674
+ "n": 220
2675
+ },
2676
+ "probability": {
2677
+ "nll": 0.804850408692942,
2678
+ "brier": 0.5061907772166189,
2679
+ "jsd": 0.22027018987749325,
2680
+ "acc": 0.5727272727272728,
2681
+ "acc_k": 0.5727272727272728,
2682
+ "ece": 0.12663529461080378,
2683
+ "sel_acc80": 0.6136363636363636,
2684
+ "coverage": 1.0,
2685
+ "false_abstain": 0.0,
2686
+ "null_recall": "n/a",
2687
+ "n": 220
2688
+ },
2689
+ "numeric": {
2690
+ "nll": 0.854408628287633,
2691
+ "brier": 0.6255809344308126,
2692
+ "jsd": 0.24148170952539819,
2693
+ "acc": 0.4909090909090909,
2694
+ "acc_k": 0.4909090909090909,
2695
+ "ece": 0.24755196327512913,
2696
+ "sel_acc80": 0.4772727272727273,
2697
+ "coverage": 1.0,
2698
+ "false_abstain": 0.0,
2699
+ "null_recall": "n/a",
2700
+ "n": 220
2701
+ },
2702
+ "tradeoff": {
2703
+ "nll": 0.9582777382892255,
2704
+ "brier": 0.6041841936847155,
2705
+ "jsd": 0.23730402818816942,
2706
+ "acc": 0.5409090909090909,
2707
+ "acc_k": 0.5409090909090909,
2708
+ "ece": 0.21417698643424293,
2709
+ "sel_acc80": 0.5965909090909091,
2710
+ "coverage": 1.0,
2711
+ "false_abstain": 0.0,
2712
+ "null_recall": "n/a",
2713
+ "n": 220
2714
+ }
2715
+ },
2716
+ "by_qtype": {},
2717
+ "by_label_class": {
2718
+ "answer": {
2719
+ "nll": 0.8425315491747045,
2720
+ "brier": 0.566066637960518,
2721
+ "jsd": 0.22709053783966102,
2722
+ "acc": 0.5443181818181818,
2723
+ "acc_k": 0.5443181818181818,
2724
+ "ece": 0.17712931768460705,
2725
+ "sel_acc80": 0.5738636363636364,
2726
+ "coverage": 1.0,
2727
+ "false_abstain": 0.0,
2728
+ "null_recall": "n/a",
2729
+ "n": 880
2730
+ }
2731
+ },
2732
+ "baselines": {
2733
+ "majority_class": 1,
2734
+ "majority_acc": 0.4602272727272727
2735
+ }
2736
+ },
2737
+ "wh_rubric_flip": {
2738
+ "n": 2206,
2739
+ "n_total": 2206,
2740
+ "suffix_overflow": 0,
2741
+ "raw": {
2742
+ "nll": 0.5215484658596726,
2743
+ "brier": 0.33112145645101976,
2744
+ "jsd": 0.13730397267005034,
2745
+ "acc": 0.742973708068903,
2746
+ "acc_k": 0.7501196744853997,
2747
+ "ece": 0.08118663341860714,
2748
+ "sel_acc80": 0.8164305949008499,
2749
+ "conf_wrong": 0.0,
2750
+ "coverage": 0.9301903898458749,
2751
+ "false_abstain": 0.022498803255146003,
2752
+ "null_recall": 0.9145299145299145,
2753
+ "auroc_null": 0.9924267530777824,
2754
+ "auroc_conf": 0.9735570530209114
2755
+ },
2756
+ "by_family": {
2757
+ "loan_approval": {
2758
+ "nll": 0.03986105987120578,
2759
+ "brier": 0.022708606037172275,
2760
+ "jsd": 0.00911794257520729,
2761
+ "acc": 0.9861111111111112,
2762
+ "acc_k": 0.9858156028368794,
2763
+ "ece": 0.014913849118683096,
2764
+ "sel_acc80": 1.0,
2765
+ "conf_wrong": 0.0,
2766
+ "coverage": 0.9791666666666666,
2767
+ "false_abstain": 0.0,
2768
+ "null_recall": 1.0,
2769
+ "auroc_null": 1.0,
2770
+ "auroc_conf": 1.0,
2771
+ "n": 144
2772
+ },
2773
+ "shift_swap": {
2774
+ "nll": 0.3556272740222601,
2775
+ "brier": 0.19030986550930082,
2776
+ "jsd": 0.07994201505668534,
2777
+ "acc": 0.8554216867469879,
2778
+ "acc_k": 0.8782051282051282,
2779
+ "ece": 0.05131704243550819,
2780
+ "sel_acc80": 0.9398496240601504,
2781
+ "conf_wrong": 0.0,
2782
+ "coverage": 0.927710843373494,
2783
+ "false_abstain": 0.02564102564102564,
2784
+ "null_recall": 0.8,
2785
+ "auroc_null": 0.9878205128205129,
2786
+ "auroc_conf": 0.948076923076923,
2787
+ "n": 166
2788
+ },
2789
+ "vendor_onboarding": {
2790
+ "nll": 0.2460892508917783,
2791
+ "brier": 0.14163160753660067,
2792
+ "jsd": 0.06893476279267717,
2793
+ "acc": 0.9102564102564102,
2794
+ "acc_k": 0.9014084507042254,
2795
+ "ece": 0.054101008826341355,
2796
+ "sel_acc80": 0.976,
2797
+ "conf_wrong": 0.0,
2798
+ "coverage": 0.9102564102564102,
2799
+ "false_abstain": 0.0,
2800
+ "null_recall": 1.0,
2801
+ "auroc_null": 1.0,
2802
+ "auroc_conf": 1.0,
2803
+ "n": 156
2804
+ },
2805
+ "content_moderation": {
2806
+ "nll": 0.5707868104784541,
2807
+ "brier": 0.3399881864130089,
2808
+ "jsd": 0.1565919344337606,
2809
+ "acc": 0.7297297297297297,
2810
+ "acc_k": 0.8166666666666667,
2811
+ "ece": 0.07247401008734834,
2812
+ "sel_acc80": 0.7983193277310925,
2813
+ "conf_wrong": 0.0,
2814
+ "coverage": 0.6891891891891893,
2815
+ "false_abstain": 0.16666666666666666,
2816
+ "null_recall": 0.9285714285714286,
2817
+ "auroc_null": 0.9428571428571428,
2818
+ "auroc_conf": 0.9383928571428571,
2819
+ "n": 148
2820
+ },
2821
+ "warranty_claim": {
2822
+ "nll": 0.1533494565143874,
2823
+ "brier": 0.08736313763136712,
2824
+ "jsd": 0.038794858019505045,
2825
+ "acc": 0.935064935064935,
2826
+ "acc_k": 0.9342105263157895,
2827
+ "ece": 0.03250992956099568,
2828
+ "sel_acc80": 1.0,
2829
+ "conf_wrong": 0.0,
2830
+ "coverage": 0.987012987012987,
2831
+ "false_abstain": 0.0,
2832
+ "null_recall": 1.0,
2833
+ "auroc_null": 1.0,
2834
+ "auroc_conf": 1.0,
2835
+ "n": 154
2836
+ },
2837
+ "visa_extension": {
2838
+ "nll": 0.3792518490833633,
2839
+ "brier": 0.22995354363423323,
2840
+ "jsd": 0.10051831683881524,
2841
+ "acc": 0.8493150684931506,
2842
+ "acc_k": 0.8790322580645161,
2843
+ "ece": 0.06351610569104751,
2844
+ "sel_acc80": 0.9230769230769231,
2845
+ "conf_wrong": 0.0,
2846
+ "coverage": 0.7945205479452055,
2847
+ "false_abstain": 0.07258064516129033,
2848
+ "null_recall": 0.9545454545454546,
2849
+ "auroc_null": 0.9736070381231671,
2850
+ "auroc_conf": 0.9721407624633431,
2851
+ "n": 146
2852
+ },
2853
+ "procurement_request": {
2854
+ "nll": 0.3881801381792412,
2855
+ "brier": 0.18757470893100156,
2856
+ "jsd": 0.07515705849986501,
2857
+ "acc": 0.8909090909090909,
2858
+ "acc_k": 0.9292929292929293,
2859
+ "ece": 0.08982028690251438,
2860
+ "sel_acc80": 0.9204545454545454,
2861
+ "conf_wrong": 0.0,
2862
+ "coverage": 0.8818181818181818,
2863
+ "false_abstain": 0.04040404040404041,
2864
+ "null_recall": 0.8181818181818182,
2865
+ "auroc_null": 0.977961432506887,
2866
+ "auroc_conf": 0.9687786960514233,
2867
+ "n": 110
2868
+ },
2869
+ "research_ethics_review": {
2870
+ "nll": 0.4421237581343384,
2871
+ "brier": 0.24773885698642706,
2872
+ "jsd": 0.11269835014231615,
2873
+ "acc": 0.8333333333333334,
2874
+ "acc_k": 0.8315789473684211,
2875
+ "ece": 0.103033944412514,
2876
+ "sel_acc80": 0.9080459770114943,
2877
+ "conf_wrong": 0.0,
2878
+ "coverage": 0.8425925925925926,
2879
+ "false_abstain": 0.042105263157894736,
2880
+ "null_recall": 1.0,
2881
+ "auroc_null": 0.9854251012145749,
2882
+ "auroc_conf": 0.9838056680161944,
2883
+ "n": 108
2884
+ },
2885
+ "subscription_cancellation": {
2886
+ "nll": 0.2644555372286351,
2887
+ "brier": 0.15680793083414146,
2888
+ "jsd": 0.06982178903755307,
2889
+ "acc": 0.8728813559322034,
2890
+ "acc_k": 0.9326923076923077,
2891
+ "ece": 0.061388740095041576,
2892
+ "sel_acc80": 0.9578947368421052,
2893
+ "conf_wrong": 0.0,
2894
+ "coverage": 0.8559322033898304,
2895
+ "false_abstain": 0.057692307692307696,
2896
+ "null_recall": 0.7857142857142857,
2897
+ "auroc_null": 0.9745879120879122,
2898
+ "auroc_conf": 0.9684065934065934,
2899
+ "n": 118
2900
+ },
2901
+ "feature_flag_rollout": {
2902
+ "nll": 0.2267314556395052,
2903
+ "brier": 0.1367964508738713,
2904
+ "jsd": 0.06290056031954905,
2905
+ "acc": 0.8913043478260869,
2906
+ "acc_k": 0.8913043478260869,
2907
+ "ece": 0.04663619204707767,
2908
+ "sel_acc80": 1.0,
2909
+ "coverage": 1.0,
2910
+ "false_abstain": 0.0,
2911
+ "null_recall": "n/a",
2912
+ "n": 46
2913
+ },
2914
+ "insurance_claim": {
2915
+ "nll": 0.33329503028518,
2916
+ "brier": 0.185439438184308,
2917
+ "jsd": 0.07467994362635967,
2918
+ "acc": 0.8333333333333334,
2919
+ "acc_k": 0.8333333333333334,
2920
+ "ece": 0.11369339625040695,
2921
+ "sel_acc80": 1.0,
2922
+ "coverage": 1.0,
2923
+ "false_abstain": 0.0,
2924
+ "null_recall": "n/a",
2925
+ "n": 30
2926
+ },
2927
+ "temporal": {
2928
+ "nll": 0.7525550379081508,
2929
+ "brier": 0.5281081163658832,
2930
+ "jsd": 0.20925588567068154,
2931
+ "acc": 0.5727272727272728,
2932
+ "acc_k": 0.5727272727272728,
2933
+ "ece": 0.1753265808929097,
2934
+ "sel_acc80": 0.625,
2935
+ "coverage": 1.0,
2936
+ "false_abstain": 0.0,
2937
+ "null_recall": "n/a",
2938
+ "n": 220
2939
+ },
2940
+ "numeric": {
2941
+ "nll": 0.8554757856422319,
2942
+ "brier": 0.6267431100109008,
2943
+ "jsd": 0.2417873556438029,
2944
+ "acc": 0.4909090909090909,
2945
+ "acc_k": 0.4909090909090909,
2946
+ "ece": 0.2482705327597531,
2947
+ "sel_acc80": 0.4659090909090909,
2948
+ "coverage": 1.0,
2949
+ "false_abstain": 0.0,
2950
+ "null_recall": "n/a",
2951
+ "n": 220
2952
+ },
2953
+ "probability": {
2954
+ "nll": 0.8047299060130065,
2955
+ "brier": 0.5061678304732362,
2956
+ "jsd": 0.220257933543653,
2957
+ "acc": 0.5727272727272728,
2958
+ "acc_k": 0.5727272727272728,
2959
+ "ece": 0.11572440903295172,
2960
+ "sel_acc80": 0.6136363636363636,
2961
+ "coverage": 1.0,
2962
+ "false_abstain": 0.0,
2963
+ "null_recall": "n/a",
2964
+ "n": 220
2965
+ },
2966
+ "tradeoff": {
2967
+ "nll": 0.9591729335220512,
2968
+ "brier": 0.6044553194922022,
2969
+ "jsd": 0.2374179176613618,
2970
+ "acc": 0.5454545454545454,
2971
+ "acc_k": 0.5454545454545454,
2972
+ "ece": 0.20948043235323643,
2973
+ "sel_acc80": 0.5909090909090909,
2974
+ "coverage": 1.0,
2975
+ "false_abstain": 0.0,
2976
+ "null_recall": "n/a",
2977
+ "n": 220
2978
+ }
2979
+ },
2980
+ "by_qtype": {},
2981
+ "by_label_class": {
2982
+ "answer": {
2983
+ "nll": 0.5344578349380356,
2984
+ "brier": 0.340925948758752,
2985
+ "jsd": 0.13998262548140297,
2986
+ "acc": 0.7333652465294399,
2987
+ "acc_k": 0.7501196744853997,
2988
+ "ece": 0.09119710230610484,
2989
+ "sel_acc80": 0.8068181818181818,
2990
+ "coverage": 0.977501196744854,
2991
+ "false_abstain": 0.022498803255146003,
2992
+ "null_recall": "n/a",
2993
+ "n": 2089
2994
+ },
2995
+ "null": {
2996
+ "nll": 0.2910555427425767,
2997
+ "brier": 0.15606517926424532,
2998
+ "jsd": 0.08947742802974543,
2999
+ "acc": 0.9145299145299145,
3000
+ "ece": 0.1208790940606696,
3001
+ "sel_acc80": 0.9680851063829787,
3002
+ "conf_wrong": 0.0,
3003
+ "coverage": 0.0854700854700855,
3004
+ "false_abstain": "n/a",
3005
+ "null_recall": 0.9145299145299145,
3006
+ "n": 117
3007
+ }
3008
+ },
3009
+ "baselines": {
3010
+ "majority_class": 0,
3011
+ "majority_acc": 0.39981867633726204
3012
+ }
3013
+ },
3014
+ "wh_rubric_shuffled": {
3015
+ "n": 2437,
3016
+ "n_total": 2437,
3017
+ "suffix_overflow": 0,
3018
+ "raw": {
3019
+ "nll": 3.3731109702781983,
3020
+ "brier": 1.0166531838701214,
3021
+ "jsd": 0.3668337750954112,
3022
+ "acc": 0.37792367665162085,
3023
+ "acc_k": 0.4059104737070839,
3024
+ "ece": 0.46684057980616855,
3025
+ "sel_acc80": 0.3553846153846154,
3026
+ "conf_wrong": 0.17647058823529413,
3027
+ "coverage": 0.926138695116947,
3028
+ "false_abstain": 0.049109083007388095,
3029
+ "null_recall": 0.49264705882352944,
3030
+ "auroc_null": 0.9573203466523506,
3031
+ "auroc_conf": 0.8061951964618963
3032
+ },
3033
+ "by_family": {
3034
+ "temporal": {
3035
+ "nll": 1.0978670096057097,
3036
+ "brier": 0.7736457181852393,
3037
+ "jsd": 0.2832470099926861,
3038
+ "acc": 0.4090909090909091,
3039
+ "acc_k": 0.4090909090909091,
3040
+ "ece": 0.34149774312973025,
3041
+ "sel_acc80": 0.3693181818181818,
3042
+ "coverage": 1.0,
3043
+ "false_abstain": 0.0,
3044
+ "null_recall": "n/a",
3045
+ "n": 220
3046
+ },
3047
+ "shift_swap": {
3048
+ "nll": 4.813958214372273,
3049
+ "brier": 1.1627913595629051,
3050
+ "jsd": 0.41357994903252204,
3051
+ "acc": 0.35602094240837695,
3052
+ "acc_k": 0.37222222222222223,
3053
+ "ece": 0.5618253638919112,
3054
+ "sel_acc80": 0.35294117647058826,
3055
+ "conf_wrong": 0.2727272727272727,
3056
+ "coverage": 0.9319371727748691,
3057
+ "false_abstain": 0.044444444444444446,
3058
+ "null_recall": 0.45454545454545453,
3059
+ "auroc_null": 0.961111111111111,
3060
+ "auroc_conf": 0.8565656565656565,
3061
+ "n": 191
3062
+ },
3063
+ "vendor_onboarding": {
3064
+ "nll": 4.468025309710248,
3065
+ "brier": 1.1531663283624074,
3066
+ "jsd": 0.4069967513654248,
3067
+ "acc": 0.31958762886597936,
3068
+ "acc_k": 0.39325842696629215,
3069
+ "ece": 0.5820424177290239,
3070
+ "sel_acc80": 0.34615384615384615,
3071
+ "conf_wrong": 0.1875,
3072
+ "coverage": 0.9175257731958762,
3073
+ "false_abstain": 0.07303370786516854,
3074
+ "null_recall": 0.1875,
3075
+ "auroc_null": 0.9371488764044944,
3076
+ "auroc_conf": 0.8433988764044944,
3077
+ "n": 194
3078
+ },
3079
+ "loan_approval": {
3080
+ "nll": 7.430180710009384,
3081
+ "brier": 1.3288401062960586,
3082
+ "jsd": 0.4609968152472018,
3083
+ "acc": 0.33146067415730335,
3084
+ "acc_k": 0.3505747126436782,
3085
+ "ece": 0.6675498291347803,
3086
+ "sel_acc80": 0.3356643356643357,
3087
+ "conf_wrong": 0.5,
3088
+ "coverage": 0.9662921348314607,
3089
+ "false_abstain": 0.022988505747126436,
3090
+ "null_recall": 0.5,
3091
+ "auroc_null": 0.978448275862069,
3092
+ "auroc_conf": 0.9396551724137931,
3093
+ "n": 178
3094
+ },
3095
+ "probability": {
3096
+ "nll": 1.5095582926439517,
3097
+ "brier": 0.9146099048716316,
3098
+ "jsd": 0.35344105098573364,
3099
+ "acc": 0.35,
3100
+ "acc_k": 0.35,
3101
+ "ece": 0.3669138718735088,
3102
+ "sel_acc80": 0.3181818181818182,
3103
+ "coverage": 1.0,
3104
+ "false_abstain": 0.0,
3105
+ "null_recall": "n/a",
3106
+ "n": 220
3107
+ },
3108
+ "feature_flag_rollout": {
3109
+ "nll": 4.963032018178103,
3110
+ "brier": 1.250659962667729,
3111
+ "jsd": 0.4386572342122217,
3112
+ "acc": 0.34782608695652173,
3113
+ "acc_k": 0.34782608695652173,
3114
+ "ece": 0.6342515031928602,
3115
+ "sel_acc80": 0.3783783783783784,
3116
+ "coverage": 1.0,
3117
+ "false_abstain": 0.0,
3118
+ "null_recall": "n/a",
3119
+ "n": 46
3120
+ },
3121
+ "research_ethics_review": {
3122
+ "nll": 3.379301129427671,
3123
+ "brier": 1.1293730274385836,
3124
+ "jsd": 0.4101792599802621,
3125
+ "acc": 0.31746031746031744,
3126
+ "acc_k": 0.38392857142857145,
3127
+ "ece": 0.5505591124769241,
3128
+ "sel_acc80": 0.3069306930693069,
3129
+ "conf_wrong": 0.07142857142857142,
3130
+ "coverage": 0.8492063492063492,
3131
+ "false_abstain": 0.125,
3132
+ "null_recall": 0.35714285714285715,
3133
+ "auroc_null": 0.8852040816326531,
3134
+ "auroc_conf": 0.7633928571428572,
3135
+ "n": 126
3136
+ },
3137
+ "warranty_claim": {
3138
+ "nll": 5.425085813770316,
3139
+ "brier": 1.222989340354703,
3140
+ "jsd": 0.42426978503284835,
3141
+ "acc": 0.35978835978835977,
3142
+ "acc_k": 0.3675675675675676,
3143
+ "ece": 0.6112382035091439,
3144
+ "sel_acc80": 0.3684210526315789,
3145
+ "conf_wrong": 0.5,
3146
+ "coverage": 0.9788359788359788,
3147
+ "false_abstain": 0.010810810810810811,
3148
+ "null_recall": 0.5,
3149
+ "auroc_null": 0.9905405405405405,
3150
+ "auroc_conf": 0.8783783783783783,
3151
+ "n": 189
3152
+ },
3153
+ "numeric": {
3154
+ "nll": 0.7009633054040385,
3155
+ "brier": 0.5083317633419103,
3156
+ "jsd": 0.19623566233246967,
3157
+ "acc": 0.5636363636363636,
3158
+ "acc_k": 0.5636363636363636,
3159
+ "ece": 0.20822090967134996,
3160
+ "sel_acc80": 0.5795454545454546,
3161
+ "coverage": 1.0,
3162
+ "false_abstain": 0.0,
3163
+ "null_recall": "n/a",
3164
+ "n": 220
3165
+ },
3166
+ "content_moderation": {
3167
+ "nll": 2.532750401458349,
3168
+ "brier": 0.9723702132508161,
3169
+ "jsd": 0.3676846987407566,
3170
+ "acc": 0.3670886075949367,
3171
+ "acc_k": 0.4094488188976378,
3172
+ "ece": 0.4052060521856139,
3173
+ "sel_acc80": 0.3464566929133858,
3174
+ "conf_wrong": 0.06451612903225806,
3175
+ "coverage": 0.6645569620253164,
3176
+ "false_abstain": 0.23622047244094488,
3177
+ "null_recall": 0.7419354838709677,
3178
+ "auroc_null": 0.8270256540513081,
3179
+ "auroc_conf": 0.7680975361950725,
3180
+ "n": 158
3181
+ },
3182
+ "visa_extension": {
3183
+ "nll": 3.4319388280589163,
3184
+ "brier": 1.185777007774496,
3185
+ "jsd": 0.4211514399505007,
3186
+ "acc": 0.3023255813953488,
3187
+ "acc_k": 0.3698630136986301,
3188
+ "ece": 0.5790469986061717,
3189
+ "sel_acc80": 0.2971014492753623,
3190
+ "conf_wrong": 0.11538461538461539,
3191
+ "coverage": 0.8023255813953488,
3192
+ "false_abstain": 0.14383561643835616,
3193
+ "null_recall": 0.5,
3194
+ "auroc_null": 0.8730242360379347,
3195
+ "auroc_conf": 0.8132244467860906,
3196
+ "n": 172
3197
+ },
3198
+ "subscription_cancellation": {
3199
+ "nll": 4.124055456988184,
3200
+ "brier": 1.142628609921072,
3201
+ "jsd": 0.40527967431266604,
3202
+ "acc": 0.3546099290780142,
3203
+ "acc_k": 0.424,
3204
+ "ece": 0.5671253876483187,
3205
+ "sel_acc80": 0.35398230088495575,
3206
+ "conf_wrong": 0.3125,
3207
+ "coverage": 0.8652482269503546,
3208
+ "false_abstain": 0.096,
3209
+ "null_recall": 0.4375,
3210
+ "auroc_null": 0.9035,
3211
+ "auroc_conf": 0.801,
3212
+ "n": 141
3213
+ },
3214
+ "procurement_request": {
3215
+ "nll": 5.064798743828172,
3216
+ "brier": 1.1244811567056938,
3217
+ "jsd": 0.3945059149254654,
3218
+ "acc": 0.37593984962406013,
3219
+ "acc_k": 0.4369747899159664,
3220
+ "ece": 0.5752717774165303,
3221
+ "sel_acc80": 0.3925233644859813,
3222
+ "conf_wrong": 0.21428571428571427,
3223
+ "coverage": 0.8796992481203008,
3224
+ "false_abstain": 0.07563025210084033,
3225
+ "null_recall": 0.5,
3226
+ "auroc_null": 0.9357743097238895,
3227
+ "auroc_conf": 0.8715486194477791,
3228
+ "n": 133
3229
+ },
3230
+ "tradeoff": {
3231
+ "nll": 1.295216810789425,
3232
+ "brier": 0.8154554789002251,
3233
+ "jsd": 0.30319613299194453,
3234
+ "acc": 0.4318181818181818,
3235
+ "acc_k": 0.4318181818181818,
3236
+ "ece": 0.35886814662001354,
3237
+ "sel_acc80": 0.3806818181818182,
3238
+ "coverage": 1.0,
3239
+ "false_abstain": 0.0,
3240
+ "null_recall": "n/a",
3241
+ "n": 220
3242
+ },
3243
+ "insurance_claim": {
3244
+ "nll": 5.987324790766054,
3245
+ "brier": 1.150424499572334,
3246
+ "jsd": 0.407765815124327,
3247
+ "acc": 0.41379310344827586,
3248
+ "acc_k": 0.41379310344827586,
3249
+ "ece": 0.5872780434016523,
3250
+ "sel_acc80": 0.4166666666666667,
3251
+ "coverage": 1.0,
3252
+ "false_abstain": 0.0,
3253
+ "null_recall": "n/a",
3254
+ "n": 29
3255
+ }
3256
+ },
3257
+ "by_qtype": {},
3258
+ "by_label_class": {
3259
+ "answer": {
3260
+ "nll": 3.4790294568471882,
3261
+ "brier": 1.0318484361922742,
3262
+ "jsd": 0.3710045446095289,
3263
+ "acc": 0.37114298131247286,
3264
+ "acc_k": 0.4059104737070839,
3265
+ "ece": 0.47994829282352375,
3266
+ "sel_acc80": 0.3508962520369365,
3267
+ "coverage": 0.9508909169926119,
3268
+ "false_abstain": 0.049109083007388095,
3269
+ "null_recall": "n/a",
3270
+ "n": 2301
3271
+ },
3272
+ "null": {
3273
+ "nll": 1.5810636350190332,
3274
+ "brier": 0.7595629221548795,
3275
+ "jsd": 0.2962680350072858,
3276
+ "acc": 0.49264705882352944,
3277
+ "ece": 0.3134835373391124,
3278
+ "sel_acc80": 0.45871559633027525,
3279
+ "conf_wrong": 0.17647058823529413,
3280
+ "coverage": 0.5073529411764706,
3281
+ "false_abstain": "n/a",
3282
+ "null_recall": 0.49264705882352944,
3283
+ "n": 136
3284
+ }
3285
+ },
3286
+ "baselines": {
3287
+ "majority_class": 0,
3288
+ "majority_acc": 0.4095199015182602
3289
+ }
3290
+ }
3291
+ }
3292
+ }
inference/example.py ADDED
@@ -0,0 +1,27 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """uv run --no-sync python inference/example.py"""
2
+ from typical import Typical
3
+
4
+ STATE = (
5
+ "Ticket #4821: Customer says their package arrived damaged. They want a replacement "
6
+ "shipped overnight, not a refund. Order was placed 3 days ago, still under warranty."
7
+ )
8
+
9
+ m = Typical.from_pretrained("OzLabs/typical-small", device="auto")
10
+
11
+ # K-way choice over a fixed label set.
12
+ print(m.choice(STATE, "What does the customer want?", ["refund", "replacement", "repair"]))
13
+
14
+ # Yes/no.
15
+ print(m.noul(STATE, "Is the order still under warranty?"))
16
+
17
+ # Ordinal levels + expected value.
18
+ print(m.score(STATE, "How urgent is this ticket?", ["0", "1", "2", "3"]))
19
+
20
+ # Full JevBench-style decide(), with the runtime block (latency, p_null, truncation flags).
21
+ probs, runtime = m.decide(
22
+ STATE,
23
+ {"type": "choice", "instructions": "What does the customer want?",
24
+ "criteria": {"refund": "money back", "replacement": "a new item shipped"}},
25
+ ["refund", "replacement"],
26
+ )
27
+ print(probs, runtime)
inference/requirements.txt ADDED
@@ -0,0 +1,5 @@
 
 
 
 
 
 
1
+ torch>=2.14.0
2
+ transformers>=5.17.0
3
+ safetensors>=0.4.0
4
+ huggingface_hub>=0.25.0
5
+ numpy>=2.0.0
inference/test_parity.py ADDED
@@ -0,0 +1,121 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Parity check: inference/typical/ vs the training repo's pcdm_jev.decider.PCDMDecider
2
+ (mode="native"), same checkpoint (guychuk/pcdm-runs/typical-small/best.pt), same device,
3
+ same 6 items. Downloads a 1.7B checkpoint -> gated like the repo's other real-checkpoint
4
+ tests (tests/test_pipeline.py, tests/test_jevbench.py): RUN_SLOW=1 required.
5
+
6
+ uv run --no-sync pytest inference/test_parity.py -v # (with RUN_SLOW=1 set)
7
+ """
8
+ import os
9
+ import sys
10
+ from pathlib import Path
11
+
12
+ import pytest
13
+
14
+ REPO_ROOT = Path(__file__).resolve().parent.parent
15
+ INFERENCE_DIR = Path(__file__).resolve().parent
16
+ for p in (str(REPO_ROOT), str(INFERENCE_DIR)):
17
+ if p not in sys.path:
18
+ sys.path.insert(0, p)
19
+
20
+ CKPT_REPO = "guychuk/pcdm-runs"
21
+ CKPT_FILE = "typical-small/best.pt"
22
+
23
+ STATE = (
24
+ "Ticket #4821: Customer says their package arrived damaged. They want a replacement "
25
+ "shipped overnight, not a refund. Order was placed 3 days ago, still under warranty."
26
+ )
27
+
28
+ ITEMS = [
29
+ ("What does the customer want?", ["refund", "replacement", "repair"]),
30
+ ("Is the order still under warranty?", ["no", "yes"]),
31
+ ("How urgent is this ticket?", ["0", "1", "2", "3"]),
32
+ ("What is the customer's sentiment?", ["positive", "neutral", "negative"]),
33
+ ("Should this be escalated to a manager?", ["no", "yes"]),
34
+ ("Which department should own this ticket?", ["billing", "shipping", "support"]),
35
+ ]
36
+
37
+ NOUL_ITEMS = [
38
+ "Is the order still under warranty?",
39
+ "Should this be escalated to a manager?",
40
+ "Was the package damaged?",
41
+ ]
42
+
43
+
44
+ @pytest.mark.skipif(not os.environ.get("RUN_SLOW"), reason="downloads the real typical-small "
45
+ "checkpoint (Qwen3-1.7B-Base + LoRA); set RUN_SLOW=1 to run")
46
+ def test_typical_matches_pcdm_decider():
47
+ from huggingface_hub import hf_hub_download
48
+
49
+ from pcdm_jev.decider import PCDMDecider
50
+ from typical import Typical, to_labels
51
+
52
+ local_best_pt = hf_hub_download(CKPT_REPO, CKPT_FILE)
53
+ run_dir = os.path.dirname(local_best_pt)
54
+
55
+ orig = PCDMDecider(run_dir=run_dir, mode="native", device="auto")
56
+ mine = Typical.from_pretrained(CKPT_REPO, filename=CKPT_FILE, device=orig.device,
57
+ max_state=orig.max_state)
58
+ assert orig.device == mine.device
59
+
60
+ max_diff = 0.0
61
+ for query, labels in ITEMS:
62
+ p_orig = orig._probs("native", STATE, query, labels)
63
+ p_mine = mine._raw(STATE, query, labels)
64
+ diff = (p_orig.float().cpu() - p_mine.float().cpu()).abs().max().item()
65
+ max_diff = max(max_diff, diff)
66
+ assert diff < 1e-4, f"{query!r} {labels}: max abs diff {diff}"
67
+ print(f"[parity] {len(ITEMS)} items, max abs prob diff = {max_diff:.2e}")
68
+
69
+ # noul reversed-label control: the checkpoint's noul_head is bern (query-only, position-
70
+ # invariant by construction -- the rendered suffix never contains the candidate text, so
71
+ # P(yes) under ["no","yes"] and ["yes","no"] is the SAME forward pass mathematically).
72
+ # ponytail: exact float equality is verified on CPU (checked separately, bit-identical);
73
+ # on MPS the two calls can differ by ~1e-7 from non-deterministic reduction order, so this
74
+ # check uses a tight epsilon rather than `==` to stay device-agnostic.
75
+ eps = 1e-6
76
+ for q in NOUL_ITEMS:
77
+ p_yes_forward = mine.noul(STATE, q)
78
+ probs_rev, _ = to_labels(mine._raw(STATE, q, ["yes", "no"]), ["yes", "no"])
79
+ assert abs(probs_rev["yes"] - p_yes_forward) < eps, (q, probs_rev["yes"], p_yes_forward)
80
+
81
+ orig_fwd, _ = to_labels(orig._probs("native", STATE, q, ["no", "yes"]), ["no", "yes"])
82
+ orig_rev, _ = to_labels(orig._probs("native", STATE, q, ["yes", "no"]), ["yes", "no"])
83
+ assert abs(orig_fwd["yes"] - orig_rev["yes"]) < eps, (q, orig_fwd["yes"], orig_rev["yes"])
84
+ assert abs(orig_fwd["yes"] - p_yes_forward) < 1e-4, (q, orig_fwd["yes"], p_yes_forward)
85
+ print(f"[parity] noul reversed-label check OK over {len(NOUL_ITEMS)} items (eps={eps})")
86
+
87
+
88
+ @pytest.mark.skipif(not os.environ.get("RUN_SLOW"), reason="downloads the real "
89
+ "typical-small-preview checkpoint; set RUN_SLOW=1 to run")
90
+ def test_typical_noul_via_choice_path_on_preview():
91
+ """typical-small-preview trains noul_head="choice" (no Bernoulli head): m.noul() must
92
+ still work by scoring the plain K-way ["no","yes"] choice and reading off P(yes) -- and
93
+ must still match the original decider exactly (order-invariance is a bern-only guarantee,
94
+ not required here)."""
95
+ from huggingface_hub import hf_hub_download
96
+
97
+ from pcdm_jev.decider import PCDMDecider
98
+ from typical import Typical, to_labels
99
+
100
+ local_best_pt = hf_hub_download(CKPT_REPO, "typical-small-preview/best.pt")
101
+ run_dir = os.path.dirname(local_best_pt)
102
+
103
+ orig = PCDMDecider(run_dir=run_dir, mode="native", device="auto")
104
+ assert orig.native["m"].noul_head == "choice"
105
+ mine = Typical.from_pretrained(CKPT_REPO, filename="typical-small-preview/best.pt",
106
+ device=orig.device, max_state=orig.max_state)
107
+ assert mine.model.noul_head == "choice"
108
+
109
+ for q in NOUL_ITEMS:
110
+ p_yes = mine.noul(STATE, q)
111
+ assert 0.0 <= p_yes <= 1.0
112
+ orig_probs, _ = to_labels(orig._probs("native", STATE, q, ["no", "yes"]), ["no", "yes"])
113
+ assert abs(orig_probs["yes"] - p_yes) < 1e-4, (q, orig_probs["yes"], p_yes)
114
+ print(f"[parity] preview noul() via K-way choice path OK over {len(NOUL_ITEMS)} items")
115
+
116
+
117
+ if __name__ == "__main__":
118
+ if not os.environ.get("RUN_SLOW"):
119
+ os.environ["RUN_SLOW"] = "1"
120
+ test_typical_matches_pcdm_decider()
121
+ test_typical_noul_via_choice_path_on_preview()
inference/typical/__init__.py ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ from .core import Typical, query_text, to_labels
2
+
3
+ __all__ = ["Typical", "query_text", "to_labels"]
inference/typical/backbone.py ADDED
@@ -0,0 +1,90 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Frozen backbone + LoRA adapter -- trimmed port of the training repo's encode.py, for
2
+ inference only. Only what native_kv_decide touches: tokenizer, .model (the raw HF causal
3
+ trunk, tap-truncated), LoRA weight loading. No FeatureCache/EmbedEncoder/forward() --
4
+ the native serving path reads hidden states straight off Backbone.model, never through a
5
+ Backbone.forward() (that method, and the frozen-feature/h_top machinery it supports, only
6
+ exists in the training repo's energy-head path). extra_tap is also dropped: both shipped
7
+ checkpoints (typical-small, typical-small-preview) train with extra_tap=0.
8
+ """
9
+ import math
10
+
11
+ import torch
12
+ import torch.nn as nn
13
+ from transformers import AutoModel, AutoTokenizer
14
+
15
+ _LORA_MODULES = ("q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj")
16
+ _printed_device = False
17
+
18
+
19
+ def pick_device(device: str = "auto") -> str:
20
+ global _printed_device
21
+ if device != "auto":
22
+ resolved = device
23
+ elif torch.cuda.is_available():
24
+ resolved = "cuda"
25
+ elif torch.backends.mps.is_available():
26
+ resolved = "mps"
27
+ else:
28
+ resolved = "cpu"
29
+ if not _printed_device:
30
+ print(f"typical: device={resolved}")
31
+ _printed_device = True
32
+ return resolved
33
+
34
+
35
+ class LoRALinear(nn.Module):
36
+ """Frozen base nn.Linear (bf16) + a low-rank fp32 adapter on top. Numerics verbatim
37
+ from encode.py's LoRALinear (scale = alpha / r); dropout is a no-op once .eval()'d."""
38
+
39
+ def __init__(self, base: nn.Linear, r: int, alpha: float, dropout: float):
40
+ super().__init__()
41
+ self.base = base
42
+ self.base.requires_grad_(False)
43
+ self.A = nn.Parameter(torch.empty(r, base.in_features, dtype=torch.float32))
44
+ nn.init.kaiming_uniform_(self.A, a=math.sqrt(5))
45
+ self.B = nn.Parameter(torch.zeros(base.out_features, r, dtype=torch.float32))
46
+ self.drop = nn.Dropout(dropout)
47
+ self.scale = alpha / r
48
+
49
+ def forward(self, x):
50
+ out = self.base(x)
51
+ lora = self.drop(x) @ self.A.to(x.dtype).T @ self.B.to(x.dtype).T * self.scale
52
+ return out + lora
53
+
54
+
55
+ class Backbone(nn.Module):
56
+ """Frozen causal-LM trunk (Qwen3-family), truncated to tap_layer, with LoRA on its top
57
+ lora_layers blocks. Base weights come fresh from `name`; the fine-tuned delta is only
58
+ ever the LoRA A/B tensors, loaded via load_lora_state_dict from the checkpoint."""
59
+
60
+ def __init__(self, name: str, lora_layers: int = 8, lora_r: int = 16, lora_alpha: float = 32,
61
+ lora_dropout: float = 0.05, device: str = "auto", tap_layer: int = 0):
62
+ super().__init__()
63
+ self.device = pick_device(device)
64
+ self.tokenizer = AutoTokenizer.from_pretrained(name, padding_side="right")
65
+ if self.tokenizer.pad_token is None:
66
+ self.tokenizer.pad_token = self.tokenizer.eos_token
67
+ self.model = AutoModel.from_pretrained(name, dtype=torch.bfloat16, attn_implementation="sdpa")
68
+ self.model.requires_grad_(False)
69
+ self.d = self.model.config.hidden_size
70
+ n_layers = self.model.config.num_hidden_layers
71
+ # tap_layer: drop every layer above T (semantics live mid-depth, not at the final,
72
+ # next-token-shaped layer of a base LM) -- see encode.py's Backbone for the long form.
73
+ if 0 < tap_layer < n_layers:
74
+ self.model.layers = self.model.layers[:tap_layer]
75
+ self.model.config.num_hidden_layers = tap_layer
76
+ n_layers = tap_layer
77
+ if lora_r > 0:
78
+ for layer in self.model.layers[-lora_layers:]:
79
+ for parent in (layer.self_attn, layer.mlp):
80
+ for mod_name in _LORA_MODULES:
81
+ if hasattr(parent, mod_name):
82
+ setattr(parent, mod_name,
83
+ LoRALinear(getattr(parent, mod_name), lora_r, lora_alpha, lora_dropout))
84
+ self.to(self.device)
85
+ self.eval()
86
+
87
+ def load_lora_state_dict(self, sd: dict):
88
+ own = dict(self.named_parameters())
89
+ for k, v in sd.items():
90
+ own[k].data.copy_(v.to(own[k].device, own[k].dtype))
inference/typical/core.py ADDED
@@ -0,0 +1,131 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Typical: minimal from_pretrained + decide API around a native-readout PCDM checkpoint.
2
+ Mirrors pcdm_jev.decider.PCDMDecider(mode="native") exactly (same query_text/to_labels,
3
+ same native_kv_decide call) -- see that module's docstring for the mapping this
4
+ reproduces. Everything the checkpoint needs to rebuild backbone+head lives in
5
+ ckpt["args"]: backbone id, tap_layer, lora_r/lora_layers, nc_head, nc_render, null,
6
+ noul_head, score_head. z-score calibration (--zscore) needs no special handling here --
7
+ its buffers (mu_h/sd_h/mu_c/sd_c) are just part of ckpt["tower"]'s state_dict.
8
+ """
9
+ import json
10
+ import time
11
+
12
+ import torch
13
+ from huggingface_hub import hf_hub_download
14
+
15
+ from .backbone import Backbone, pick_device
16
+ from .native import NativeHead, native_kv_decide
17
+
18
+ MAX_QUERY = 64 # trained query length (see pcdm_jev.decider) -- informational only below
19
+ MAX_STATE = 256 # trained state length -- native_kv_decide's own max_state default is separate
20
+
21
+
22
+ class _Head:
23
+ """Duck-types the training repo's mcq.MCQHead just enough for native_kv_decide: a
24
+ .backbone (with .tokenizer/.model) and .device. The real MCQHead also builds a
25
+ restricted LM head over single-token letters for the mcq_zero_shot/energy readout
26
+ paths -- native_kv_decide never touches it, so it's not ported here."""
27
+
28
+ def __init__(self, backbone: Backbone):
29
+ self.backbone = backbone
30
+ self.device = backbone.device
31
+
32
+
33
+ def query_text(question: dict) -> str:
34
+ """instructions + criteria: `yes: .. no: ..` (noul), `0: .. 1: ..` (score), `label: desc`
35
+ (choice). Verbatim port of pcdm_jev.decider.query_text."""
36
+ q, crit = question["instructions"], question.get("criteria")
37
+ if not crit:
38
+ return q
39
+ if question["type"] == "noul":
40
+ rubric = f"yes: {crit.get('true', 'Yes')} no: {crit.get('false', 'No')}"
41
+ elif isinstance(crit, list):
42
+ rubric = " ".join(f"{i}: {c}" for i, c in enumerate(crit))
43
+ else:
44
+ rubric = " ".join(f"{k}: {v or k}" for k, v in crit.items())
45
+ return f"{q}\n{rubric}"
46
+
47
+
48
+ def to_labels(p: torch.Tensor, labels: list[str]) -> tuple[dict, float]:
49
+ """[K+1] (null last) -> (probs renormalised over labels, p_null). Verbatim port of
50
+ pcdm_jev.decider.to_labels."""
51
+ p = p.detach().float().cpu()
52
+ if not torch.isfinite(p).all():
53
+ raise RuntimeError("non-finite model output")
54
+ s = p[:-1].sum()
55
+ lab = p[:-1] / s if s > 0 else torch.zeros_like(p[:-1])
56
+ return dict(zip(labels, lab.tolist())), float(p[-1])
57
+
58
+
59
+ class Typical:
60
+ """from_pretrained("OzLabs/typical-small") -> a ready-to-decide native-readout model."""
61
+
62
+ def __init__(self, head: _Head, model: NativeHead, device: str, max_state: int = 4096):
63
+ self.head, self.model, self.device, self.max_state = head, model, device, max_state
64
+ self.tok = head.backbone.tokenizer
65
+
66
+ @classmethod
67
+ def from_pretrained(cls, repo_id: str, device: str = "auto", filename: str = "best.pt",
68
+ revision: str | None = None, cache_dir: str | None = None,
69
+ max_state: int = 4096) -> "Typical":
70
+ path = hf_hub_download(repo_id, filename, revision=revision, cache_dir=cache_dir)
71
+ ckpt = torch.load(path, map_location="cpu", weights_only=True)
72
+ args = ckpt.get("args", {})
73
+ if args.get("readout") != "native":
74
+ raise ValueError(f"{repo_id}/{filename} is not a native-readout checkpoint (readout={args.get('readout')!r})")
75
+ dev = pick_device(device)
76
+ backbone = Backbone(args["backbone"], lora_layers=args.get("lora_layers", 8),
77
+ lora_r=args.get("lora_r", 16), device=dev, tap_layer=args.get("tap_layer", 0))
78
+ backbone.load_lora_state_dict(ckpt["lora"])
79
+ model = NativeHead(backbone.d, nc_head=args.get("nc_head", "n2n3"), null=args.get("null", "factored"),
80
+ render=args.get("nc_render", "letters"), score_head=args.get("score_head", "choice"),
81
+ noul_head=args.get("noul_head", "choice")).to(dev)
82
+ model.load_state_dict(ckpt["tower"])
83
+ model.eval()
84
+ return cls(_Head(backbone), model, dev, max_state=max_state)
85
+
86
+ def _raw(self, state, query: str, labels: list[str]) -> torch.Tensor:
87
+ if not isinstance(state, str):
88
+ state = json.dumps(state, ensure_ascii=False)
89
+ with torch.inference_mode():
90
+ return native_kv_decide(self.head, self.model, state, [(query, labels)], max_state=self.max_state)[0]
91
+
92
+ def choice(self, state, question: str, labels: list[str]) -> dict:
93
+ """-> {label: p, ...} + "p_null"."""
94
+ probs, p_null = to_labels(self._raw(state, question, labels), labels)
95
+ return {**probs, "p_null": p_null}
96
+
97
+ def noul(self, state, question: str) -> float:
98
+ """P(yes), via the Bernoulli path if the checkpoint has one (noul_head="bern"), else
99
+ the plain K-way choice over ["no", "yes"] -- both paths are label-order invariant for
100
+ a genuine yes/no row (see native._is_bern_row / _yes_idx)."""
101
+ probs, _ = to_labels(self._raw(state, question, ["no", "yes"]), ["no", "yes"])
102
+ return probs["yes"]
103
+
104
+ def score(self, state, question: str, levels: list[str]) -> dict:
105
+ """-> {level: p, ...} + "p_null" + "expected" (E[index] under the candidate-
106
+ conditional distribution, matching the training repo's metrics.ordinal_metrics --
107
+ candidate index IS the ordinal level, by construction of how levels are rendered)."""
108
+ probs, p_null = to_labels(self._raw(state, question, levels), levels)
109
+ expected = sum(i * probs[lv] for i, lv in enumerate(levels))
110
+ return {**probs, "p_null": p_null, "expected": expected}
111
+
112
+ def decide(self, state, question: dict, labels: list[str]) -> tuple[dict, dict]:
113
+ """-> (probs over exactly `labels`, runtime block). Mirrors
114
+ pcdm_jev.decider.PCDMDecider.decide for a JevBench-style question dict
115
+ ({"type", "instructions", "criteria"})."""
116
+ if not isinstance(state, str):
117
+ state = json.dumps(state, ensure_ascii=False)
118
+ query = query_text(question)
119
+ n_state = len(self.tok(state, add_special_tokens=False)["input_ids"])
120
+ n_query = len(self.tok(query, add_special_tokens=False)["input_ids"])
121
+ t0 = time.perf_counter()
122
+ p = self._raw(state, query, labels)
123
+ probs, p_null = to_labels(p, labels)
124
+ latency = time.perf_counter() - t0
125
+ return probs, {"mode": "native", "device": self.device, "latency_s": latency, "p_null": p_null,
126
+ "state_tokens": n_state, "max_state_tokens": self.max_state,
127
+ "state_truncated": n_state > self.max_state,
128
+ "trained_max_state": MAX_STATE, "state_beyond_train_len": n_state > MAX_STATE,
129
+ "query_tokens": n_query, "query_truncated": False,
130
+ "trained_max_query": MAX_QUERY, "query_beyond_train_len": False,
131
+ "probability_origin": "native-softmax"}
inference/typical/native.py ADDED
@@ -0,0 +1,331 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """native_choice_v1/v2/v3 decision head + the KV-cached serving path -- trimmed port of the
2
+ training repo's native.py + the two mcq.py helpers it needs (_label/_ids/_pack), for
3
+ inference only. Kept: NativeHead (factored null, per-row Bernoulli "noul" routing via
4
+ _is_bern_row, the letters/tags/letters_nonull/query_only renderers) and native_kv_decide,
5
+ because pcdm_jev.decider.PCDMDecider(mode="native") calls exactly those two things and
6
+ nothing else. Dropped: run_batch_native/native_features/_score_chunked/_fit_chunks (the
7
+ training/eval batch path -- decider.py never calls them) and the n2 (Qwen3-Embedding
8
+ candidate-vector) branch, since neither shipped checkpoint (typical-small,
9
+ typical-small-preview) trains with nc_head n2/n2n3.
10
+ """
11
+ import bisect
12
+ import copy
13
+ import string
14
+
15
+ import torch
16
+ import torch.nn as nn
17
+ import torch.nn.functional as F
18
+
19
+ EMBED_DIM = 1024 # Qwen3-Embedding-0.6B hidden size; only referenced by the (unsupported) n2 branch
20
+ MAX_SUFFIX = 1024
21
+ CHOICE_OPEN, CHOICE_CLOSE = "<choice>\n", "\n</choice>\n"
22
+ LETTERS = string.ascii_uppercase + string.ascii_lowercase
23
+
24
+
25
+ def _label(i: int) -> str:
26
+ return LETTERS[i] if i < len(LETTERS) else str(i + 1)
27
+
28
+
29
+ def _render(query: str, cand_texts: list[str]):
30
+ """native_v1 (--nc_render letters): query, lettered options, a rendered "none of the
31
+ above" line, "Answer:". spans[k] = (start, end) of option k's text; spans[-1] is the
32
+ null line's text."""
33
+ text, spans = query + "\n", []
34
+ for i, c in enumerate(cand_texts + ["none of the above"]):
35
+ text += f"{_label(i)}. "
36
+ spans.append((len(text), len(text) + len(c)))
37
+ text += c + "\n"
38
+ return text + "Answer:", spans
39
+
40
+
41
+ def _render_tags(query: str, cand_texts: list[str]):
42
+ """native_v2 (--nc_render tags): letter-free, one <choice>...</choice> block per option,
43
+ no null line, no "Answer:". len(spans) == K (the null head reads a learned constant,
44
+ not a rendered line)."""
45
+ text, spans = query + "\n", []
46
+ for c in cand_texts:
47
+ text += CHOICE_OPEN
48
+ spans.append((len(text), len(text) + len(c)))
49
+ text += c + CHOICE_CLOSE
50
+ return text, spans
51
+
52
+
53
+ def _render_letters_nonull(query: str, cand_texts: list[str]):
54
+ """native_v3 (--nc_render letters_nonull): lettered options (keeps slot identity) but no
55
+ rendered null line and no "Answer:" -- both shipped checkpoints train with this render."""
56
+ text, spans = query + "\n", []
57
+ for i, c in enumerate(cand_texts):
58
+ text += f"{_label(i)}. "
59
+ spans.append((len(text), len(text) + len(c)))
60
+ text += c + "\n"
61
+ return text, spans
62
+
63
+
64
+ def _render_query_only(query: str, cand_texts: list[str]):
65
+ """--noul_head bern: suffix is the query alone (the Bernoulli head reads h_D only, never
66
+ the candidate text -- that's what makes it label-order invariant). Empty spans."""
67
+ return query + "\n", []
68
+
69
+
70
+ RENDERS = {"letters": _render, "tags": _render_tags, "letters_nonull": _render_letters_nonull,
71
+ "query_only": _render_query_only}
72
+
73
+
74
+ def _is_bern_row(cand_texts, meta=None):
75
+ """A row is bern-eligible iff its candidate set is exactly {"yes","no"} (case-insensitive),
76
+ or meta.qtype == "noul" and K == 2."""
77
+ if {c.strip().lower() for c in cand_texts} == {"yes", "no"}:
78
+ return True
79
+ return bool(meta) and meta.get("qtype") == "noul" and len(cand_texts) == 2
80
+
81
+
82
+ def _yes_idx(cand_lists, dev):
83
+ """[B] index of the literal "yes" candidate per row (case-insensitive) -- places P(yes)
84
+ there regardless of rendered order, so ["yes","no"] and ["no","yes"] give identical
85
+ P(yes). No literal "yes" -> defaults to index 0 (out-of-domain K-way row)."""
86
+ idx = []
87
+ for cl in cand_lists:
88
+ matches = [i for i, c in enumerate(cl) if c.strip().lower() == "yes"]
89
+ idx.append(matches[0] if matches else 0)
90
+ return torch.tensor(idx, dtype=torch.long, device=dev)
91
+
92
+
93
+ def _ids(tok, texts: list[str], max_len: int, offsets: bool = False):
94
+ """Token ids (no specials, truncated), optionally with char offsets; empty text -> "."."""
95
+ texts = [t if t.strip() else "." for t in texts]
96
+ enc = tok(texts, truncation=True, max_length=max_len, add_special_tokens=False, return_offsets_mapping=offsets)
97
+ return (enc["input_ids"], enc["offset_mapping"]) if offsets else enc["input_ids"]
98
+
99
+
100
+ def _pack(tok, s_ids, x_ids, tail=(), sink=True):
101
+ """rows = [eos] + state + suffix + tail, right-padded -> (input_ids, attention_mask,
102
+ lengths); sink=False drops the leading eos (native_kv_decide: the sink lives in the
103
+ cached prefix)."""
104
+ eos, pad = tok.eos_token_id, tok.pad_token_id
105
+ rows = [([eos] if sink else []) + s + x + list(tail) for s, x in zip(s_ids, x_ids)]
106
+ lengths = torch.tensor([len(r) for r in rows], dtype=torch.long)
107
+ T = max(len(r) for r in rows)
108
+ input_ids = torch.full((len(rows), T), pad, dtype=torch.long)
109
+ attention_mask = torch.zeros(len(rows), T, dtype=torch.long)
110
+ for i, r in enumerate(rows):
111
+ input_ids[i, :len(r)] = torch.tensor(r, dtype=torch.long)
112
+ attention_mask[i, :len(r)] = 1
113
+ return input_ids, attention_mask, lengths
114
+
115
+
116
+ def factored_null_logits(gate, s, c, h, cmask, temperature: float = 1.0):
117
+ """r = sigmoid(gate(z)): z is O(K) permutation-invariant set statistics over the valid
118
+ candidate scores s + rep h + masked candidate-set mean/var of c. Composed as logits so
119
+ softmax(logits) gives P(a_j) = (1-r)*p_j and P(null) = r exactly. Verbatim port of
120
+ model.factored_null_logits (self-contained: only torch/F, no other model.py state)."""
121
+ K = cmask.sum(-1)
122
+ Kf = K.float().clamp_min(1)
123
+ top2 = torch.topk(s, k=min(2, s.shape[1]), dim=-1).values
124
+ margin = top2[:, 0] - top2[:, 1] if top2.shape[1] == 2 else torch.zeros_like(top2[:, 0])
125
+ margin = torch.where(K > 1, margin, torch.zeros_like(margin))
126
+ mean_s = s.masked_fill(~cmask, 0.0).sum(-1) / Kf
127
+ lse = torch.logsumexp(s, dim=-1) - torch.log(Kf)
128
+ c0 = c.masked_fill(~cmask.unsqueeze(-1), 0.0)
129
+ mean_c = c0.sum(1) / Kf.unsqueeze(-1)
130
+ var_c = (c0 - mean_c.unsqueeze(1)).pow(2).masked_fill(~cmask.unsqueeze(-1), 0.0).sum(1) / Kf.unsqueeze(-1)
131
+ z = torch.cat([top2[:, :1], margin.unsqueeze(-1), mean_s.unsqueeze(-1), lse.unsqueeze(-1),
132
+ h, mean_c, var_c], dim=-1)
133
+ r_logit = gate(z).squeeze(-1)
134
+ log_1mr, log_r = -F.softplus(r_logit), -F.softplus(-r_logit)
135
+ s_t = (s.masked_fill(~cmask, 0.0) / temperature).masked_fill(~cmask, torch.finfo(s.dtype).min)
136
+ logp = F.log_softmax(s_t, dim=-1)
137
+ cand_logits = (logp + log_1mr.unsqueeze(-1)).masked_fill(~cmask, torch.finfo(s.dtype).min)
138
+ return torch.cat([cand_logits, log_r.unsqueeze(-1)], dim=-1)
139
+
140
+
141
+ class NativeHead(nn.Module):
142
+ """Verbatim port of native.NativeHead. n2 (nc_head in {"n2","n2n3"}) is accepted for
143
+ state_dict shape compatibility but native_kv_decide below raises if a loaded checkpoint
144
+ actually needs it -- see module docstring."""
145
+
146
+ def __init__(self, d: int, nc_head: str = "n2n3", null: str = "factored", d_proj: int = 256,
147
+ d_emb: int = EMBED_DIM, render: str = "letters", score_head: str = "choice",
148
+ noul_head: str = "choice"):
149
+ super().__init__()
150
+ assert nc_head in ("n2", "n3", "n2n3"), nc_head
151
+ assert null in ("softmax", "factored"), null
152
+ assert render in RENDERS, render
153
+ assert score_head in ("choice", "cumlink"), score_head
154
+ assert noul_head in ("choice", "bern"), noul_head
155
+ self.nc_head, self.null, self.render = nc_head, null, render
156
+ self.score_head, self.noul_head = score_head, noul_head
157
+ self.use2, self.use3 = "n2" in nc_head, "n3" in nc_head
158
+ dv = d_proj * (self.use2 + self.use3)
159
+ self.dv = dv
160
+ self.register_buffer("mu_h", torch.zeros(d)); self.register_buffer("sd_h", torch.ones(d))
161
+ self.register_buffer("mu_c", torch.zeros(d)); self.register_buffer("sd_c", torch.ones(d))
162
+ self.proj_h = nn.Sequential(nn.LayerNorm(d), nn.Linear(d, dv))
163
+ if self.use2:
164
+ self.proj_2 = nn.Sequential(nn.LayerNorm(d_emb), nn.Linear(d_emb, d_proj))
165
+ self.null_emb = nn.Parameter(torch.randn(d_emb) * 0.02)
166
+ if self.use3:
167
+ self.proj_3 = nn.Sequential(nn.LayerNorm(d), nn.Linear(d, d_proj))
168
+ self.w = nn.Linear(dv, 1)
169
+ self.res = nn.Sequential(nn.Linear(dv, dv), nn.GELU(), nn.Linear(dv, 1))
170
+ if null == "factored":
171
+ self.null_gate = nn.Sequential(nn.Linear(4 + 4 * dv, 256), nn.GELU(), nn.Linear(256, 1))
172
+ else:
173
+ self.score_null = nn.Sequential(nn.Linear(2 * dv, dv // 2), nn.GELU(), nn.Linear(dv // 2, 1))
174
+ if score_head == "cumlink" or noul_head == "bern":
175
+ self.null_gate_typed = nn.Sequential(nn.Linear(4 + 4 * d, 256), nn.GELU(), nn.Linear(256, 1))
176
+ if score_head == "cumlink":
177
+ self.w_o = nn.Linear(d, 1)
178
+ self.w_t = nn.Linear(d, 1)
179
+ self.theta0 = nn.Parameter(torch.zeros(1))
180
+ if noul_head == "bern":
181
+ self.w_n = nn.Linear(d, 1)
182
+ self.temperature = 1.0
183
+
184
+ def forward(self, h, cmask, c3=None, c2=None, yes_idx=None, bern_mask=None):
185
+ """h [B, d]; cmask [B, Kmax]; c3 [B, Kmax+1, d] (null line last); c2 [B, Kmax, d_emb].
186
+ -> logits [B, Kmax+1], null last, pads finfo.min. Verbatim port of native.NativeHead.forward."""
187
+ hz = (h - self.mu_h) / self.sd_h
188
+ if self.score_head == "cumlink":
189
+ return self._typed_logits(self._cumlink_probs(hz, c3, cmask), hz, c3, cmask)
190
+ bm = None
191
+ if self.noul_head == "bern":
192
+ bm = bern_mask if bern_mask is not None else torch.ones(h.shape[0], dtype=torch.bool, device=h.device)
193
+ if bm.all():
194
+ return self._typed_logits(self._bern_probs(hz, cmask, yes_idx), hz, c3, cmask)
195
+
196
+ B = h.shape[0]
197
+ u = self.proj_h(hz)
198
+ vs = []
199
+ if self.use2:
200
+ vs.append(self.proj_2(torch.cat([c2, self.null_emb.expand(B, 1, -1)], 1)))
201
+ if self.use3:
202
+ vs.append(self.proj_3((c3 - self.mu_c) / self.sd_c))
203
+ v = torch.cat(vs, -1)
204
+ vc, vn = v[:, :-1], v[:, -1]
205
+ uv = u.unsqueeze(1) * vc
206
+ s = uv.sum(-1) * self.dv ** -0.5 + self.w(uv).squeeze(-1) + self.res(uv).squeeze(-1)
207
+ s = s.masked_fill(~cmask, torch.finfo(s.dtype).min)
208
+ hn = torch.cat([u, vn], -1)
209
+ if self.null == "factored":
210
+ choice_logits = factored_null_logits(self.null_gate, s, vc, hn, cmask, self.temperature)
211
+ else:
212
+ choice_logits = torch.cat([s, self.score_null(hn)], dim=-1)
213
+
214
+ if bm is not None and bm.any():
215
+ bern_logits = self._typed_logits(self._bern_probs(hz, cmask, yes_idx), hz, c3, cmask)
216
+ return torch.where(bm.unsqueeze(-1), bern_logits, choice_logits)
217
+ return choice_logits
218
+
219
+ def _bern_probs(self, hz, cmask, yes_idx):
220
+ B, Kmax = cmask.shape
221
+ p_yes = torch.sigmoid(self.w_n(hz).squeeze(-1))
222
+ idx = yes_idx if yes_idx is not None else torch.ones(B, dtype=torch.long, device=hz.device)
223
+ K = cmask.sum(-1)
224
+ other = (1 - p_yes) / (K - 1).clamp_min(1).float()
225
+ p = other.unsqueeze(-1).expand(B, Kmax) * cmask.float()
226
+ ar = torch.arange(B, device=hz.device)
227
+ p[ar, idx] = p_yes
228
+ return p.masked_fill(~cmask, 0.0)
229
+
230
+ def _cumlink_probs(self, hz, c3, cmask):
231
+ B, Kmax = cmask.shape
232
+ c3z = (c3 - self.mu_c) / self.sd_c
233
+ u = self.w_o(hz).squeeze(-1)
234
+ width = F.softplus(self.w_t(c3z[:, :-1]).squeeze(-1)).masked_fill(~cmask, 0.0)
235
+ theta = self.theta0 + torch.cumsum(width, dim=-1)
236
+ cdf = torch.sigmoid(theta - u.unsqueeze(-1))
237
+ p_prev = F.pad(cdf, (1, 0))[:, :-1]
238
+ p = (cdf - p_prev).clamp_min(0.0)
239
+ K = cmask.sum(-1)
240
+ last = (K - 1).clamp_min(0)
241
+ remainder = torch.where(K > 1, 1.0 - cdf.gather(1, (last - 1).clamp_min(0).unsqueeze(1)).squeeze(1),
242
+ torch.ones_like(u))
243
+ p = p.scatter(1, last.unsqueeze(1), remainder.unsqueeze(1)).masked_fill(~cmask, 0.0)
244
+ return p
245
+
246
+ def _typed_logits(self, p, hz, c3, cmask):
247
+ c3z = (c3 - self.mu_c) / self.sd_c
248
+ s = torch.log(p.clamp_min(1e-12)).masked_fill(~cmask, torch.finfo(p.dtype).min)
249
+ hn = torch.cat([hz, c3z[:, -1]], dim=-1)
250
+ return factored_null_logits(self.null_gate_typed, s, c3z[:, :-1], hn, cmask, self.temperature)
251
+
252
+
253
+ def _pool_matrix(spans, offsets, base, Kmax, T, K=None):
254
+ """[Kmax+1, T] mean-pooling rows over each option's rendered char span (bisected against
255
+ the tokenizer's char offsets)."""
256
+ K = len(spans) - 1 if K is None else K
257
+ starts, ends = [o[0] for o in offsets], [o[1] for o in offsets]
258
+ pool = torch.zeros(Kmax + 1, T)
259
+ for k, (cs, ce) in enumerate(spans):
260
+ j0, j1 = bisect.bisect_right(ends, cs), bisect.bisect_left(starts, ce)
261
+ pool[Kmax if k == K else k, base + j0:base + j1] = 1.0
262
+ return pool / pool.sum(-1, keepdim=True).clamp_min(1.0)
263
+
264
+
265
+ def _causal_pad_mask(attn2d, q_len, dtype):
266
+ """Explicit additive (batch,1,q_len,kv_len) mask -- transformers' automatic 2D-mask +
267
+ past_key_values path mis-handles a right-padded batch sharing one KV cache; this
268
+ reproduces the un-batched per-row computation exactly."""
269
+ batch, kv_len = attn2d.shape
270
+ past_len = kv_len - q_len
271
+ q_idx = torch.arange(q_len, device=attn2d.device).unsqueeze(1)
272
+ kv_idx = torch.arange(kv_len, device=attn2d.device).unsqueeze(0)
273
+ causal = kv_idx <= (past_len + q_idx)
274
+ allowed = causal.unsqueeze(0) & attn2d.bool().unsqueeze(1)
275
+ mask = torch.zeros(batch, 1, q_len, kv_len, dtype=dtype, device=attn2d.device)
276
+ return mask.masked_fill(~allowed.unsqueeze(1), torch.finfo(dtype).min)
277
+
278
+
279
+ @torch.inference_mode()
280
+ def native_kv_decide(head, model, state, queries, chunk: int = 32, max_state: int = 256,
281
+ max_suffix: int = MAX_SUFFIX, vec_cache=None):
282
+ """Serving path: [eos] + state encoded ONCE into a KV cache on head.backbone.model; every
283
+ query's suffix runs against that cached prefix, `chunk` queries per forward. queries =
284
+ [(query, candidates)] -> list of [K_i + 1] probability vectors (null last). Verbatim port
285
+ of native.native_kv_decide (head/model contract unchanged -- head needs only .backbone
286
+ (tokenizer + model) and .device; model is a NativeHead). No hierarchical fallback: a
287
+ suffix longer than max_suffix raises (raise max_suffix instead)."""
288
+ if model.use2:
289
+ raise NotImplementedError(
290
+ "nc_head n2/n2n3 (Qwen3-Embedding candidate vectors) isn't ported to the "
291
+ "inference package -- neither typical-small nor typical-small-preview needs it. "
292
+ "Port encode.EmbedEncoder + native._n2_vectors from the training repo if a future "
293
+ "checkpoint requires it."
294
+ )
295
+ tok, dev, lm = head.backbone.tokenizer, head.device, head.backbone.model
296
+ prefix = torch.tensor([[tok.eos_token_id] + _ids(tok, [state], max_state)[0]], device=dev)
297
+ state_cache = lm(input_ids=prefix, use_cache=True).past_key_values
298
+ Ls = prefix.shape[1]
299
+ out = []
300
+ for start in range(0, len(queries), chunk):
301
+ qs = queries[start:start + chunk]
302
+ m = len(qs)
303
+ is_bern = [model.noul_head == "bern" and _is_bern_row(c) for _, c in qs]
304
+ rendered = [RENDERS["query_only" if b else getattr(model, "render", "letters")](q, c)
305
+ for b, (q, c) in zip(is_bern, qs)]
306
+ enc = tok([r[0] for r in rendered], add_special_tokens=False, return_offsets_mapping=True)
307
+ x_ids, offs = enc["input_ids"], enc["offset_mapping"]
308
+ assert max(len(x) for x in x_ids) <= max_suffix, f"suffix > max_suffix={max_suffix}; raise it"
309
+ input_ids, am, lengths = _pack(tok, [[]] * m, x_ids, tail=[tok.eos_token_id], sink=False)
310
+ T = input_ids.shape[1]
311
+ Kmax = max(len(c) for _, c in qs)
312
+ pool = torch.stack([_pool_matrix(r[1], o, 0, Kmax, T, K=len(c)) for r, o, (_, c) in zip(rendered, offs, qs)])
313
+ cmask = torch.zeros(m, Kmax, dtype=torch.bool)
314
+ for i, (_, c) in enumerate(qs):
315
+ cmask[i, :len(c)] = True
316
+
317
+ cache = copy.deepcopy(state_cache)
318
+ cache.batch_repeat_interleave(m)
319
+ attn = torch.cat([torch.ones(m, Ls, dtype=torch.long), am], dim=1).to(dev)
320
+ position_ids = (torch.arange(T) + Ls).expand(m, -1).to(dev)
321
+ H = lm(input_ids=input_ids.to(dev), attention_mask=_causal_pad_mask(attn, T, lm.dtype),
322
+ position_ids=position_ids, past_key_values=cache, use_cache=False).last_hidden_state.float()
323
+ h = H[torch.arange(m, device=dev), (lengths - 1).to(dev)]
324
+ C3 = torch.bmm(pool.to(dev), H)
325
+ C2 = None # n2 unsupported -- guarded above
326
+ yi = _yes_idx([c for _, c in qs], dev) if model.noul_head == "bern" else None
327
+ bm = torch.tensor(is_bern, dtype=torch.bool, device=dev) if model.noul_head == "bern" else None
328
+ probs = torch.softmax(model(h, cmask.to(dev), C3, C2, yes_idx=yi, bern_mask=bm), dim=-1)
329
+ for i, (_, c) in enumerate(qs):
330
+ out.append(torch.cat([probs[i, :len(c)], probs[i, -1:]]))
331
+ return out
jevbench_summary.json ADDED
@@ -0,0 +1,2824 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "easy": {
3
+ "accuracy": 1.0,
4
+ "brier_mean": 0.011964783954900882,
5
+ "calibration_n": 48,
6
+ "complete": true,
7
+ "cost_basis": [
8
+ "local_gpu_no_provider_tariff"
9
+ ],
10
+ "coverage": 1.0,
11
+ "dataset_hash": "3cbd0e951fa8059741955ef2bd3273b0bb60aecae8f4a68b89872f55f7376b3d",
12
+ "ece": {
13
+ "bins": [
14
+ {
15
+ "accuracy": null,
16
+ "hi": 0.1,
17
+ "lo": 0.0,
18
+ "mean_confidence": null,
19
+ "n": 0
20
+ },
21
+ {
22
+ "accuracy": null,
23
+ "hi": 0.2,
24
+ "lo": 0.1,
25
+ "mean_confidence": null,
26
+ "n": 0
27
+ },
28
+ {
29
+ "accuracy": null,
30
+ "hi": 0.3,
31
+ "lo": 0.2,
32
+ "mean_confidence": null,
33
+ "n": 0
34
+ },
35
+ {
36
+ "accuracy": null,
37
+ "hi": 0.4,
38
+ "lo": 0.3,
39
+ "mean_confidence": null,
40
+ "n": 0
41
+ },
42
+ {
43
+ "accuracy": null,
44
+ "hi": 0.5,
45
+ "lo": 0.4,
46
+ "mean_confidence": null,
47
+ "n": 0
48
+ },
49
+ {
50
+ "accuracy": 1.0,
51
+ "hi": 0.6,
52
+ "lo": 0.5,
53
+ "mean_confidence": 0.5003303289413452,
54
+ "n": 1
55
+ },
56
+ {
57
+ "accuracy": null,
58
+ "hi": 0.7,
59
+ "lo": 0.6,
60
+ "mean_confidence": null,
61
+ "n": 0
62
+ },
63
+ {
64
+ "accuracy": null,
65
+ "hi": 0.8,
66
+ "lo": 0.7,
67
+ "mean_confidence": null,
68
+ "n": 0
69
+ },
70
+ {
71
+ "accuracy": 1.0,
72
+ "hi": 0.9,
73
+ "lo": 0.8,
74
+ "mean_confidence": 0.8724372982978821,
75
+ "n": 2
76
+ },
77
+ {
78
+ "accuracy": 1.0,
79
+ "hi": 1.0,
80
+ "lo": 0.9,
81
+ "mean_confidence": 0.9944215284453498,
82
+ "n": 45
83
+ }
84
+ ],
85
+ "ece": 0.020954714467128106,
86
+ "n": 48
87
+ },
88
+ "latency": {
89
+ "n": 48,
90
+ "p50_s": 0.06915471702814102,
91
+ "p95_s": 0.10234953109174966
92
+ },
93
+ "latency_failures": {
94
+ "n": 0,
95
+ "p50_s": null,
96
+ "p95_s": null
97
+ },
98
+ "ledger_charged_usd": null,
99
+ "macro_accuracy": 1.0,
100
+ "model_identities": [
101
+ "pcdm:native:runs/tm1b"
102
+ ],
103
+ "n_attempted": 48,
104
+ "n_correct": 48,
105
+ "n_planned": 48,
106
+ "n_renormalized": 0,
107
+ "n_scorable": 48,
108
+ "n_valid": 48,
109
+ "operational_success": 1.0,
110
+ "ordinal_mae": null,
111
+ "paraphrase_consistency": {
112
+ "agree": 0,
113
+ "agreement": null,
114
+ "both_correct_rate_all_pairs": null,
115
+ "both_valid": 0,
116
+ "pairs": 0
117
+ },
118
+ "per_family": {
119
+ "extraction": {
120
+ "accuracy": 1.0,
121
+ "brier_mean": 0.003787310148891915,
122
+ "calibration_n": 12,
123
+ "cost_basis": [
124
+ "local_gpu_no_provider_tariff"
125
+ ],
126
+ "coverage": 1.0,
127
+ "ece": {
128
+ "bins": [
129
+ {
130
+ "accuracy": null,
131
+ "hi": 0.1,
132
+ "lo": 0.0,
133
+ "mean_confidence": null,
134
+ "n": 0
135
+ },
136
+ {
137
+ "accuracy": null,
138
+ "hi": 0.2,
139
+ "lo": 0.1,
140
+ "mean_confidence": null,
141
+ "n": 0
142
+ },
143
+ {
144
+ "accuracy": null,
145
+ "hi": 0.3,
146
+ "lo": 0.2,
147
+ "mean_confidence": null,
148
+ "n": 0
149
+ },
150
+ {
151
+ "accuracy": null,
152
+ "hi": 0.4,
153
+ "lo": 0.3,
154
+ "mean_confidence": null,
155
+ "n": 0
156
+ },
157
+ {
158
+ "accuracy": null,
159
+ "hi": 0.5,
160
+ "lo": 0.4,
161
+ "mean_confidence": null,
162
+ "n": 0
163
+ },
164
+ {
165
+ "accuracy": null,
166
+ "hi": 0.6,
167
+ "lo": 0.5,
168
+ "mean_confidence": null,
169
+ "n": 0
170
+ },
171
+ {
172
+ "accuracy": null,
173
+ "hi": 0.7,
174
+ "lo": 0.6,
175
+ "mean_confidence": null,
176
+ "n": 0
177
+ },
178
+ {
179
+ "accuracy": null,
180
+ "hi": 0.8,
181
+ "lo": 0.7,
182
+ "mean_confidence": null,
183
+ "n": 0
184
+ },
185
+ {
186
+ "accuracy": 1.0,
187
+ "hi": 0.9,
188
+ "lo": 0.8,
189
+ "mean_confidence": 0.8558686375617981,
190
+ "n": 1
191
+ },
192
+ {
193
+ "accuracy": 1.0,
194
+ "hi": 1.0,
195
+ "lo": 0.9,
196
+ "mean_confidence": 0.9874209653247487,
197
+ "n": 11
198
+ }
199
+ ],
200
+ "ece": 0.023541728655497224,
201
+ "n": 12
202
+ },
203
+ "latency": {
204
+ "n": 12,
205
+ "p50_s": 0.06931246258318424,
206
+ "p95_s": 0.08583159018307922
207
+ },
208
+ "latency_failures": {
209
+ "n": 0,
210
+ "p50_s": null,
211
+ "p95_s": null
212
+ },
213
+ "n_attempted": 12,
214
+ "n_correct": 12,
215
+ "n_planned": 12,
216
+ "n_renormalized": 0,
217
+ "n_scorable": 12,
218
+ "n_valid": 12,
219
+ "operational_success": 1.0,
220
+ "ordinal_mae": null,
221
+ "price_per_1000_decisions_usd": null,
222
+ "schema_validity": 1.0,
223
+ "schema_validity_strict": 1.0
224
+ },
225
+ "fact": {
226
+ "accuracy": 1.0,
227
+ "brier_mean": 0.044071809342586504,
228
+ "calibration_n": 12,
229
+ "cost_basis": [
230
+ "local_gpu_no_provider_tariff"
231
+ ],
232
+ "coverage": 1.0,
233
+ "ece": {
234
+ "bins": [
235
+ {
236
+ "accuracy": null,
237
+ "hi": 0.1,
238
+ "lo": 0.0,
239
+ "mean_confidence": null,
240
+ "n": 0
241
+ },
242
+ {
243
+ "accuracy": null,
244
+ "hi": 0.2,
245
+ "lo": 0.1,
246
+ "mean_confidence": null,
247
+ "n": 0
248
+ },
249
+ {
250
+ "accuracy": null,
251
+ "hi": 0.3,
252
+ "lo": 0.2,
253
+ "mean_confidence": null,
254
+ "n": 0
255
+ },
256
+ {
257
+ "accuracy": null,
258
+ "hi": 0.4,
259
+ "lo": 0.3,
260
+ "mean_confidence": null,
261
+ "n": 0
262
+ },
263
+ {
264
+ "accuracy": null,
265
+ "hi": 0.5,
266
+ "lo": 0.4,
267
+ "mean_confidence": null,
268
+ "n": 0
269
+ },
270
+ {
271
+ "accuracy": 1.0,
272
+ "hi": 0.6,
273
+ "lo": 0.5,
274
+ "mean_confidence": 0.5003303289413452,
275
+ "n": 1
276
+ },
277
+ {
278
+ "accuracy": null,
279
+ "hi": 0.7,
280
+ "lo": 0.6,
281
+ "mean_confidence": null,
282
+ "n": 0
283
+ },
284
+ {
285
+ "accuracy": null,
286
+ "hi": 0.8,
287
+ "lo": 0.7,
288
+ "mean_confidence": null,
289
+ "n": 0
290
+ },
291
+ {
292
+ "accuracy": 1.0,
293
+ "hi": 0.9,
294
+ "lo": 0.8,
295
+ "mean_confidence": 0.8890059590339661,
296
+ "n": 1
297
+ },
298
+ {
299
+ "accuracy": 1.0,
300
+ "hi": 1.0,
301
+ "lo": 0.9,
302
+ "mean_confidence": 0.988783085346222,
303
+ "n": 10
304
+ }
305
+ ],
306
+ "ece": 0.06023607154687243,
307
+ "n": 12
308
+ },
309
+ "latency": {
310
+ "n": 12,
311
+ "p50_s": 0.07128027640283108,
312
+ "p95_s": 0.1037625202909112
313
+ },
314
+ "latency_failures": {
315
+ "n": 0,
316
+ "p50_s": null,
317
+ "p95_s": null
318
+ },
319
+ "n_attempted": 12,
320
+ "n_correct": 12,
321
+ "n_planned": 12,
322
+ "n_renormalized": 0,
323
+ "n_scorable": 12,
324
+ "n_valid": 12,
325
+ "operational_success": 1.0,
326
+ "ordinal_mae": null,
327
+ "price_per_1000_decisions_usd": null,
328
+ "schema_validity": 1.0,
329
+ "schema_validity_strict": 1.0
330
+ },
331
+ "intent": {
332
+ "accuracy": 1.0,
333
+ "brier_mean": 1.4921552066029873e-08,
334
+ "calibration_n": 12,
335
+ "cost_basis": [
336
+ "local_gpu_no_provider_tariff"
337
+ ],
338
+ "coverage": 1.0,
339
+ "ece": {
340
+ "bins": [
341
+ {
342
+ "accuracy": null,
343
+ "hi": 0.1,
344
+ "lo": 0.0,
345
+ "mean_confidence": null,
346
+ "n": 0
347
+ },
348
+ {
349
+ "accuracy": null,
350
+ "hi": 0.2,
351
+ "lo": 0.1,
352
+ "mean_confidence": null,
353
+ "n": 0
354
+ },
355
+ {
356
+ "accuracy": null,
357
+ "hi": 0.3,
358
+ "lo": 0.2,
359
+ "mean_confidence": null,
360
+ "n": 0
361
+ },
362
+ {
363
+ "accuracy": null,
364
+ "hi": 0.4,
365
+ "lo": 0.3,
366
+ "mean_confidence": null,
367
+ "n": 0
368
+ },
369
+ {
370
+ "accuracy": null,
371
+ "hi": 0.5,
372
+ "lo": 0.4,
373
+ "mean_confidence": null,
374
+ "n": 0
375
+ },
376
+ {
377
+ "accuracy": null,
378
+ "hi": 0.6,
379
+ "lo": 0.5,
380
+ "mean_confidence": null,
381
+ "n": 0
382
+ },
383
+ {
384
+ "accuracy": null,
385
+ "hi": 0.7,
386
+ "lo": 0.6,
387
+ "mean_confidence": null,
388
+ "n": 0
389
+ },
390
+ {
391
+ "accuracy": null,
392
+ "hi": 0.8,
393
+ "lo": 0.7,
394
+ "mean_confidence": null,
395
+ "n": 0
396
+ },
397
+ {
398
+ "accuracy": null,
399
+ "hi": 0.9,
400
+ "lo": 0.8,
401
+ "mean_confidence": null,
402
+ "n": 0
403
+ },
404
+ {
405
+ "accuracy": 1.0,
406
+ "hi": 1.0,
407
+ "lo": 0.9,
408
+ "mean_confidence": 0.9999667853116989,
409
+ "n": 12
410
+ }
411
+ ],
412
+ "ece": 3.3214688301086426e-05,
413
+ "n": 12
414
+ },
415
+ "latency": {
416
+ "n": 12,
417
+ "p50_s": 0.0686907097697258,
418
+ "p95_s": 0.1342240288853645
419
+ },
420
+ "latency_failures": {
421
+ "n": 0,
422
+ "p50_s": null,
423
+ "p95_s": null
424
+ },
425
+ "n_attempted": 12,
426
+ "n_correct": 12,
427
+ "n_planned": 12,
428
+ "n_renormalized": 0,
429
+ "n_scorable": 12,
430
+ "n_valid": 12,
431
+ "operational_success": 1.0,
432
+ "ordinal_mae": null,
433
+ "price_per_1000_decisions_usd": null,
434
+ "schema_validity": 1.0,
435
+ "schema_validity_strict": 1.0
436
+ },
437
+ "tool_selection": {
438
+ "accuracy": 1.0,
439
+ "brier_mean": 1.4065730446045023e-09,
440
+ "calibration_n": 12,
441
+ "cost_basis": [
442
+ "local_gpu_no_provider_tariff"
443
+ ],
444
+ "coverage": 1.0,
445
+ "ece": {
446
+ "bins": [
447
+ {
448
+ "accuracy": null,
449
+ "hi": 0.1,
450
+ "lo": 0.0,
451
+ "mean_confidence": null,
452
+ "n": 0
453
+ },
454
+ {
455
+ "accuracy": null,
456
+ "hi": 0.2,
457
+ "lo": 0.1,
458
+ "mean_confidence": null,
459
+ "n": 0
460
+ },
461
+ {
462
+ "accuracy": null,
463
+ "hi": 0.3,
464
+ "lo": 0.2,
465
+ "mean_confidence": null,
466
+ "n": 0
467
+ },
468
+ {
469
+ "accuracy": null,
470
+ "hi": 0.4,
471
+ "lo": 0.3,
472
+ "mean_confidence": null,
473
+ "n": 0
474
+ },
475
+ {
476
+ "accuracy": null,
477
+ "hi": 0.5,
478
+ "lo": 0.4,
479
+ "mean_confidence": null,
480
+ "n": 0
481
+ },
482
+ {
483
+ "accuracy": null,
484
+ "hi": 0.6,
485
+ "lo": 0.5,
486
+ "mean_confidence": null,
487
+ "n": 0
488
+ },
489
+ {
490
+ "accuracy": null,
491
+ "hi": 0.7,
492
+ "lo": 0.6,
493
+ "mean_confidence": null,
494
+ "n": 0
495
+ },
496
+ {
497
+ "accuracy": null,
498
+ "hi": 0.8,
499
+ "lo": 0.7,
500
+ "mean_confidence": null,
501
+ "n": 0
502
+ },
503
+ {
504
+ "accuracy": null,
505
+ "hi": 0.9,
506
+ "lo": 0.8,
507
+ "mean_confidence": null,
508
+ "n": 0
509
+ },
510
+ {
511
+ "accuracy": 1.0,
512
+ "hi": 1.0,
513
+ "lo": 0.9,
514
+ "mean_confidence": 0.9999921570221583,
515
+ "n": 12
516
+ }
517
+ ],
518
+ "ece": 7.842977841732157e-06,
519
+ "n": 12
520
+ },
521
+ "latency": {
522
+ "n": 12,
523
+ "p50_s": 0.06903873011469841,
524
+ "p95_s": 0.07324958574026823
525
+ },
526
+ "latency_failures": {
527
+ "n": 0,
528
+ "p50_s": null,
529
+ "p95_s": null
530
+ },
531
+ "n_attempted": 12,
532
+ "n_correct": 12,
533
+ "n_planned": 12,
534
+ "n_renormalized": 0,
535
+ "n_scorable": 12,
536
+ "n_valid": 12,
537
+ "operational_success": 1.0,
538
+ "ordinal_mae": null,
539
+ "price_per_1000_decisions_usd": null,
540
+ "schema_validity": 1.0,
541
+ "schema_validity_strict": 1.0
542
+ }
543
+ },
544
+ "price_per_1000_decisions_usd": null,
545
+ "probability_sources": [
546
+ "native"
547
+ ],
548
+ "schema_validity": 1.0,
549
+ "schema_validity_strict": 1.0,
550
+ "splits": {
551
+ "public": {
552
+ "accuracy": 1.0,
553
+ "brier_mean": 0.011964783954900882,
554
+ "calibration_n": 48,
555
+ "cost_basis": [
556
+ "local_gpu_no_provider_tariff"
557
+ ],
558
+ "coverage": 1.0,
559
+ "ece": {
560
+ "bins": [
561
+ {
562
+ "accuracy": null,
563
+ "hi": 0.1,
564
+ "lo": 0.0,
565
+ "mean_confidence": null,
566
+ "n": 0
567
+ },
568
+ {
569
+ "accuracy": null,
570
+ "hi": 0.2,
571
+ "lo": 0.1,
572
+ "mean_confidence": null,
573
+ "n": 0
574
+ },
575
+ {
576
+ "accuracy": null,
577
+ "hi": 0.3,
578
+ "lo": 0.2,
579
+ "mean_confidence": null,
580
+ "n": 0
581
+ },
582
+ {
583
+ "accuracy": null,
584
+ "hi": 0.4,
585
+ "lo": 0.3,
586
+ "mean_confidence": null,
587
+ "n": 0
588
+ },
589
+ {
590
+ "accuracy": null,
591
+ "hi": 0.5,
592
+ "lo": 0.4,
593
+ "mean_confidence": null,
594
+ "n": 0
595
+ },
596
+ {
597
+ "accuracy": 1.0,
598
+ "hi": 0.6,
599
+ "lo": 0.5,
600
+ "mean_confidence": 0.5003303289413452,
601
+ "n": 1
602
+ },
603
+ {
604
+ "accuracy": null,
605
+ "hi": 0.7,
606
+ "lo": 0.6,
607
+ "mean_confidence": null,
608
+ "n": 0
609
+ },
610
+ {
611
+ "accuracy": null,
612
+ "hi": 0.8,
613
+ "lo": 0.7,
614
+ "mean_confidence": null,
615
+ "n": 0
616
+ },
617
+ {
618
+ "accuracy": 1.0,
619
+ "hi": 0.9,
620
+ "lo": 0.8,
621
+ "mean_confidence": 0.8724372982978821,
622
+ "n": 2
623
+ },
624
+ {
625
+ "accuracy": 1.0,
626
+ "hi": 1.0,
627
+ "lo": 0.9,
628
+ "mean_confidence": 0.9944215284453498,
629
+ "n": 45
630
+ }
631
+ ],
632
+ "ece": 0.020954714467128106,
633
+ "n": 48
634
+ },
635
+ "latency": {
636
+ "n": 48,
637
+ "p50_s": 0.06915471702814102,
638
+ "p95_s": 0.10234953109174966
639
+ },
640
+ "latency_failures": {
641
+ "n": 0,
642
+ "p50_s": null,
643
+ "p95_s": null
644
+ },
645
+ "n_attempted": 48,
646
+ "n_correct": 48,
647
+ "n_planned": 48,
648
+ "n_renormalized": 0,
649
+ "n_scorable": 48,
650
+ "n_valid": 48,
651
+ "operational_success": 1.0,
652
+ "ordinal_mae": null,
653
+ "price_per_1000_decisions_usd": null,
654
+ "schema_validity": 1.0,
655
+ "schema_validity_strict": 1.0
656
+ }
657
+ },
658
+ "version": "jevbench-v1"
659
+ },
660
+ "hard": {
661
+ "accuracy": 0.42342342342342343,
662
+ "brier_mean": 0.7731682475813886,
663
+ "calibration_n": 111,
664
+ "complete": true,
665
+ "cost_basis": [
666
+ "local_gpu_no_provider_tariff"
667
+ ],
668
+ "coverage": 1.0,
669
+ "dataset_hash": "e816e106f562dec75a8da12e69b75e957a4bbd9f6300f1145fc71837327a3c61",
670
+ "ece": {
671
+ "bins": [
672
+ {
673
+ "accuracy": null,
674
+ "hi": 0.1,
675
+ "lo": 0.0,
676
+ "mean_confidence": null,
677
+ "n": 0
678
+ },
679
+ {
680
+ "accuracy": null,
681
+ "hi": 0.2,
682
+ "lo": 0.1,
683
+ "mean_confidence": null,
684
+ "n": 0
685
+ },
686
+ {
687
+ "accuracy": 0.0,
688
+ "hi": 0.3,
689
+ "lo": 0.2,
690
+ "mean_confidence": 0.27849190682172775,
691
+ "n": 4
692
+ },
693
+ {
694
+ "accuracy": 0.125,
695
+ "hi": 0.4,
696
+ "lo": 0.3,
697
+ "mean_confidence": 0.37296687439084053,
698
+ "n": 8
699
+ },
700
+ {
701
+ "accuracy": 0.4166666666666667,
702
+ "hi": 0.5,
703
+ "lo": 0.4,
704
+ "mean_confidence": 0.4450460250178973,
705
+ "n": 12
706
+ },
707
+ {
708
+ "accuracy": 0.35,
709
+ "hi": 0.6,
710
+ "lo": 0.5,
711
+ "mean_confidence": 0.5447314709424973,
712
+ "n": 20
713
+ },
714
+ {
715
+ "accuracy": 0.5,
716
+ "hi": 0.7,
717
+ "lo": 0.6,
718
+ "mean_confidence": 0.6454204533781324,
719
+ "n": 14
720
+ },
721
+ {
722
+ "accuracy": 0.3333333333333333,
723
+ "hi": 0.8,
724
+ "lo": 0.7,
725
+ "mean_confidence": 0.7522401312987009,
726
+ "n": 6
727
+ },
728
+ {
729
+ "accuracy": 0.25,
730
+ "hi": 0.9,
731
+ "lo": 0.8,
732
+ "mean_confidence": 0.8441648408770561,
733
+ "n": 8
734
+ },
735
+ {
736
+ "accuracy": 0.5897435897435898,
737
+ "hi": 1.0,
738
+ "lo": 0.9,
739
+ "mean_confidence": 0.965268883949671,
740
+ "n": 39
741
+ }
742
+ ],
743
+ "ece": 0.2818109610596219,
744
+ "n": 111
745
+ },
746
+ "latency": {
747
+ "n": 111,
748
+ "p50_s": 0.11237071454524994,
749
+ "p95_s": 0.14899961836636066
750
+ },
751
+ "latency_failures": {
752
+ "n": 0,
753
+ "p50_s": null,
754
+ "p95_s": null
755
+ },
756
+ "ledger_charged_usd": null,
757
+ "macro_accuracy": 0.4814144429701705,
758
+ "model_identities": [
759
+ "pcdm:native:runs/tm1b"
760
+ ],
761
+ "n_attempted": 111,
762
+ "n_correct": 47,
763
+ "n_planned": 111,
764
+ "n_renormalized": 0,
765
+ "n_scorable": 111,
766
+ "n_valid": 111,
767
+ "operational_success": 1.0,
768
+ "ordinal_mae": 0.7110752028723558,
769
+ "paraphrase_consistency": {
770
+ "agree": 0,
771
+ "agreement": null,
772
+ "both_correct_rate_all_pairs": null,
773
+ "both_valid": 0,
774
+ "pairs": 0
775
+ },
776
+ "per_family": {
777
+ "adversarial": {
778
+ "accuracy": 0.6666666666666666,
779
+ "brier_mean": 0.5843859976304459,
780
+ "calibration_n": 6,
781
+ "cost_basis": [
782
+ "local_gpu_no_provider_tariff"
783
+ ],
784
+ "coverage": 1.0,
785
+ "ece": {
786
+ "bins": [
787
+ {
788
+ "accuracy": null,
789
+ "hi": 0.1,
790
+ "lo": 0.0,
791
+ "mean_confidence": null,
792
+ "n": 0
793
+ },
794
+ {
795
+ "accuracy": null,
796
+ "hi": 0.2,
797
+ "lo": 0.1,
798
+ "mean_confidence": null,
799
+ "n": 0
800
+ },
801
+ {
802
+ "accuracy": null,
803
+ "hi": 0.3,
804
+ "lo": 0.2,
805
+ "mean_confidence": null,
806
+ "n": 0
807
+ },
808
+ {
809
+ "accuracy": null,
810
+ "hi": 0.4,
811
+ "lo": 0.3,
812
+ "mean_confidence": null,
813
+ "n": 0
814
+ },
815
+ {
816
+ "accuracy": null,
817
+ "hi": 0.5,
818
+ "lo": 0.4,
819
+ "mean_confidence": null,
820
+ "n": 0
821
+ },
822
+ {
823
+ "accuracy": null,
824
+ "hi": 0.6,
825
+ "lo": 0.5,
826
+ "mean_confidence": null,
827
+ "n": 0
828
+ },
829
+ {
830
+ "accuracy": 1.0,
831
+ "hi": 0.7,
832
+ "lo": 0.6,
833
+ "mean_confidence": 0.6800556182861328,
834
+ "n": 1
835
+ },
836
+ {
837
+ "accuracy": null,
838
+ "hi": 0.8,
839
+ "lo": 0.7,
840
+ "mean_confidence": null,
841
+ "n": 0
842
+ },
843
+ {
844
+ "accuracy": 0.5,
845
+ "hi": 0.9,
846
+ "lo": 0.8,
847
+ "mean_confidence": 0.8503317534923553,
848
+ "n": 2
849
+ },
850
+ {
851
+ "accuracy": 0.6666666666666666,
852
+ "hi": 1.0,
853
+ "lo": 0.9,
854
+ "mean_confidence": 0.9747096498807272,
855
+ "n": 3
856
+ }
857
+ ],
858
+ "ece": 0.3241228063901266,
859
+ "n": 6
860
+ },
861
+ "latency": {
862
+ "n": 6,
863
+ "p50_s": 0.10217797197401524,
864
+ "p95_s": 0.10633919388055801
865
+ },
866
+ "latency_failures": {
867
+ "n": 0,
868
+ "p50_s": null,
869
+ "p95_s": null
870
+ },
871
+ "n_attempted": 6,
872
+ "n_correct": 4,
873
+ "n_planned": 6,
874
+ "n_renormalized": 0,
875
+ "n_scorable": 6,
876
+ "n_valid": 6,
877
+ "operational_success": 1.0,
878
+ "ordinal_mae": 0.0494021475315094,
879
+ "price_per_1000_decisions_usd": null,
880
+ "schema_validity": 1.0,
881
+ "schema_validity_strict": 1.0
882
+ },
883
+ "ambiguous": {
884
+ "accuracy": 0.5714285714285714,
885
+ "brier_mean": 0.6891428318952786,
886
+ "calibration_n": 7,
887
+ "cost_basis": [
888
+ "local_gpu_no_provider_tariff"
889
+ ],
890
+ "coverage": 1.0,
891
+ "ece": {
892
+ "bins": [
893
+ {
894
+ "accuracy": null,
895
+ "hi": 0.1,
896
+ "lo": 0.0,
897
+ "mean_confidence": null,
898
+ "n": 0
899
+ },
900
+ {
901
+ "accuracy": null,
902
+ "hi": 0.2,
903
+ "lo": 0.1,
904
+ "mean_confidence": null,
905
+ "n": 0
906
+ },
907
+ {
908
+ "accuracy": null,
909
+ "hi": 0.3,
910
+ "lo": 0.2,
911
+ "mean_confidence": null,
912
+ "n": 0
913
+ },
914
+ {
915
+ "accuracy": null,
916
+ "hi": 0.4,
917
+ "lo": 0.3,
918
+ "mean_confidence": null,
919
+ "n": 0
920
+ },
921
+ {
922
+ "accuracy": null,
923
+ "hi": 0.5,
924
+ "lo": 0.4,
925
+ "mean_confidence": null,
926
+ "n": 0
927
+ },
928
+ {
929
+ "accuracy": 0.5,
930
+ "hi": 0.6,
931
+ "lo": 0.5,
932
+ "mean_confidence": 0.5438312590122223,
933
+ "n": 2
934
+ },
935
+ {
936
+ "accuracy": null,
937
+ "hi": 0.7,
938
+ "lo": 0.6,
939
+ "mean_confidence": null,
940
+ "n": 0
941
+ },
942
+ {
943
+ "accuracy": 0.5,
944
+ "hi": 0.8,
945
+ "lo": 0.7,
946
+ "mean_confidence": 0.7460008561611176,
947
+ "n": 2
948
+ },
949
+ {
950
+ "accuracy": 1.0,
951
+ "hi": 0.9,
952
+ "lo": 0.8,
953
+ "mean_confidence": 0.8554508686065674,
954
+ "n": 1
955
+ },
956
+ {
957
+ "accuracy": 0.5,
958
+ "hi": 1.0,
959
+ "lo": 0.9,
960
+ "mean_confidence": 0.9503115713596344,
961
+ "n": 2
962
+ }
963
+ ],
964
+ "ece": 0.23211950063705444,
965
+ "n": 7
966
+ },
967
+ "latency": {
968
+ "n": 7,
969
+ "p50_s": 0.108931515365839,
970
+ "p95_s": 0.11300073452293874
971
+ },
972
+ "latency_failures": {
973
+ "n": 0,
974
+ "p50_s": null,
975
+ "p95_s": null
976
+ },
977
+ "n_attempted": 7,
978
+ "n_correct": 4,
979
+ "n_planned": 7,
980
+ "n_renormalized": 0,
981
+ "n_scorable": 7,
982
+ "n_valid": 7,
983
+ "operational_success": 1.0,
984
+ "ordinal_mae": null,
985
+ "price_per_1000_decisions_usd": null,
986
+ "schema_validity": 1.0,
987
+ "schema_validity_strict": 1.0
988
+ },
989
+ "judge_hard": {
990
+ "accuracy": 0.5294117647058824,
991
+ "brier_mean": 0.8882525997132734,
992
+ "calibration_n": 17,
993
+ "cost_basis": [
994
+ "local_gpu_no_provider_tariff"
995
+ ],
996
+ "coverage": 1.0,
997
+ "ece": {
998
+ "bins": [
999
+ {
1000
+ "accuracy": null,
1001
+ "hi": 0.1,
1002
+ "lo": 0.0,
1003
+ "mean_confidence": null,
1004
+ "n": 0
1005
+ },
1006
+ {
1007
+ "accuracy": null,
1008
+ "hi": 0.2,
1009
+ "lo": 0.1,
1010
+ "mean_confidence": null,
1011
+ "n": 0
1012
+ },
1013
+ {
1014
+ "accuracy": null,
1015
+ "hi": 0.3,
1016
+ "lo": 0.2,
1017
+ "mean_confidence": null,
1018
+ "n": 0
1019
+ },
1020
+ {
1021
+ "accuracy": null,
1022
+ "hi": 0.4,
1023
+ "lo": 0.3,
1024
+ "mean_confidence": null,
1025
+ "n": 0
1026
+ },
1027
+ {
1028
+ "accuracy": null,
1029
+ "hi": 0.5,
1030
+ "lo": 0.4,
1031
+ "mean_confidence": null,
1032
+ "n": 0
1033
+ },
1034
+ {
1035
+ "accuracy": null,
1036
+ "hi": 0.6,
1037
+ "lo": 0.5,
1038
+ "mean_confidence": null,
1039
+ "n": 0
1040
+ },
1041
+ {
1042
+ "accuracy": null,
1043
+ "hi": 0.7,
1044
+ "lo": 0.6,
1045
+ "mean_confidence": null,
1046
+ "n": 0
1047
+ },
1048
+ {
1049
+ "accuracy": 1.0,
1050
+ "hi": 0.8,
1051
+ "lo": 0.7,
1052
+ "mean_confidence": 0.7447604537010193,
1053
+ "n": 1
1054
+ },
1055
+ {
1056
+ "accuracy": 0.0,
1057
+ "hi": 0.9,
1058
+ "lo": 0.8,
1059
+ "mean_confidence": 0.8594815731048584,
1060
+ "n": 1
1061
+ },
1062
+ {
1063
+ "accuracy": 0.5333333333333333,
1064
+ "hi": 1.0,
1065
+ "lo": 0.9,
1066
+ "mean_confidence": 0.9754487832387289,
1067
+ "n": 15
1068
+ }
1069
+ ],
1070
+ "ece": 0.4556736981167513,
1071
+ "n": 17
1072
+ },
1073
+ "latency": {
1074
+ "n": 17,
1075
+ "p50_s": 0.1022290475666523,
1076
+ "p95_s": 0.11644351556897163
1077
+ },
1078
+ "latency_failures": {
1079
+ "n": 0,
1080
+ "p50_s": null,
1081
+ "p95_s": null
1082
+ },
1083
+ "n_attempted": 17,
1084
+ "n_correct": 9,
1085
+ "n_planned": 17,
1086
+ "n_renormalized": 0,
1087
+ "n_scorable": 17,
1088
+ "n_valid": 17,
1089
+ "operational_success": 1.0,
1090
+ "ordinal_mae": null,
1091
+ "price_per_1000_decisions_usd": null,
1092
+ "schema_validity": 1.0,
1093
+ "schema_validity_strict": 1.0
1094
+ },
1095
+ "long_policy": {
1096
+ "accuracy": 0.21052631578947367,
1097
+ "brier_mean": 0.956868090190627,
1098
+ "calibration_n": 19,
1099
+ "cost_basis": [
1100
+ "local_gpu_no_provider_tariff"
1101
+ ],
1102
+ "coverage": 1.0,
1103
+ "ece": {
1104
+ "bins": [
1105
+ {
1106
+ "accuracy": null,
1107
+ "hi": 0.1,
1108
+ "lo": 0.0,
1109
+ "mean_confidence": null,
1110
+ "n": 0
1111
+ },
1112
+ {
1113
+ "accuracy": null,
1114
+ "hi": 0.2,
1115
+ "lo": 0.1,
1116
+ "mean_confidence": null,
1117
+ "n": 0
1118
+ },
1119
+ {
1120
+ "accuracy": 0.0,
1121
+ "hi": 0.3,
1122
+ "lo": 0.2,
1123
+ "mean_confidence": 0.2781612128019333,
1124
+ "n": 2
1125
+ },
1126
+ {
1127
+ "accuracy": 0.0,
1128
+ "hi": 0.4,
1129
+ "lo": 0.3,
1130
+ "mean_confidence": 0.3873806794484456,
1131
+ "n": 3
1132
+ },
1133
+ {
1134
+ "accuracy": 0.5,
1135
+ "hi": 0.5,
1136
+ "lo": 0.4,
1137
+ "mean_confidence": 0.43897129595279694,
1138
+ "n": 4
1139
+ },
1140
+ {
1141
+ "accuracy": 0.0,
1142
+ "hi": 0.6,
1143
+ "lo": 0.5,
1144
+ "mean_confidence": 0.5172954201698303,
1145
+ "n": 1
1146
+ },
1147
+ {
1148
+ "accuracy": 0.6666666666666666,
1149
+ "hi": 0.7,
1150
+ "lo": 0.6,
1151
+ "mean_confidence": 0.6477566162745158,
1152
+ "n": 3
1153
+ },
1154
+ {
1155
+ "accuracy": 0.0,
1156
+ "hi": 0.8,
1157
+ "lo": 0.7,
1158
+ "mean_confidence": 0.7533348798751831,
1159
+ "n": 2
1160
+ },
1161
+ {
1162
+ "accuracy": 0.0,
1163
+ "hi": 0.9,
1164
+ "lo": 0.8,
1165
+ "mean_confidence": 0.8167205850283304,
1166
+ "n": 3
1167
+ },
1168
+ {
1169
+ "accuracy": 0.0,
1170
+ "hi": 1.0,
1171
+ "lo": 0.9,
1172
+ "mean_confidence": 0.9532382488250732,
1173
+ "n": 1
1174
+ }
1175
+ ],
1176
+ "ece": 0.3919302429023541,
1177
+ "n": 19
1178
+ },
1179
+ "latency": {
1180
+ "n": 19,
1181
+ "p50_s": 0.14529041200876236,
1182
+ "p95_s": 0.16292529068887227
1183
+ },
1184
+ "latency_failures": {
1185
+ "n": 0,
1186
+ "p50_s": null,
1187
+ "p95_s": null
1188
+ },
1189
+ "n_attempted": 19,
1190
+ "n_correct": 4,
1191
+ "n_planned": 19,
1192
+ "n_renormalized": 0,
1193
+ "n_scorable": 19,
1194
+ "n_valid": 19,
1195
+ "operational_success": 1.0,
1196
+ "ordinal_mae": 0.5590988993644714,
1197
+ "price_per_1000_decisions_usd": null,
1198
+ "schema_validity": 1.0,
1199
+ "schema_validity_strict": 1.0
1200
+ },
1201
+ "multi_hop": {
1202
+ "accuracy": 0.4444444444444444,
1203
+ "brier_mean": 0.6872533730651412,
1204
+ "calibration_n": 18,
1205
+ "cost_basis": [
1206
+ "local_gpu_no_provider_tariff"
1207
+ ],
1208
+ "coverage": 1.0,
1209
+ "ece": {
1210
+ "bins": [
1211
+ {
1212
+ "accuracy": null,
1213
+ "hi": 0.1,
1214
+ "lo": 0.0,
1215
+ "mean_confidence": null,
1216
+ "n": 0
1217
+ },
1218
+ {
1219
+ "accuracy": null,
1220
+ "hi": 0.2,
1221
+ "lo": 0.1,
1222
+ "mean_confidence": null,
1223
+ "n": 0
1224
+ },
1225
+ {
1226
+ "accuracy": null,
1227
+ "hi": 0.3,
1228
+ "lo": 0.2,
1229
+ "mean_confidence": null,
1230
+ "n": 0
1231
+ },
1232
+ {
1233
+ "accuracy": 1.0,
1234
+ "hi": 0.4,
1235
+ "lo": 0.3,
1236
+ "mean_confidence": 0.33783966302871704,
1237
+ "n": 1
1238
+ },
1239
+ {
1240
+ "accuracy": 0.4,
1241
+ "hi": 0.5,
1242
+ "lo": 0.4,
1243
+ "mean_confidence": 0.44226173162460325,
1244
+ "n": 5
1245
+ },
1246
+ {
1247
+ "accuracy": 0.5,
1248
+ "hi": 0.6,
1249
+ "lo": 0.5,
1250
+ "mean_confidence": 0.5598814710974693,
1251
+ "n": 8
1252
+ },
1253
+ {
1254
+ "accuracy": 0.0,
1255
+ "hi": 0.7,
1256
+ "lo": 0.6,
1257
+ "mean_confidence": 0.6466758847236633,
1258
+ "n": 2
1259
+ },
1260
+ {
1261
+ "accuracy": 0.0,
1262
+ "hi": 0.8,
1263
+ "lo": 0.7,
1264
+ "mean_confidence": 0.7700088620185852,
1265
+ "n": 1
1266
+ },
1267
+ {
1268
+ "accuracy": null,
1269
+ "hi": 0.9,
1270
+ "lo": 0.8,
1271
+ "mean_confidence": null,
1272
+ "n": 0
1273
+ },
1274
+ {
1275
+ "accuracy": 1.0,
1276
+ "hi": 1.0,
1277
+ "lo": 0.9,
1278
+ "mean_confidence": 0.9851491451263428,
1279
+ "n": 1
1280
+ }
1281
+ ],
1282
+ "ece": 0.19059623612297905,
1283
+ "n": 18
1284
+ },
1285
+ "latency": {
1286
+ "n": 18,
1287
+ "p50_s": 0.128933347761631,
1288
+ "p95_s": 0.13938846364617347
1289
+ },
1290
+ "latency_failures": {
1291
+ "n": 0,
1292
+ "p50_s": null,
1293
+ "p95_s": null
1294
+ },
1295
+ "n_attempted": 18,
1296
+ "n_correct": 8,
1297
+ "n_planned": 18,
1298
+ "n_renormalized": 0,
1299
+ "n_scorable": 18,
1300
+ "n_valid": 18,
1301
+ "operational_success": 1.0,
1302
+ "ordinal_mae": 0.4032743573188782,
1303
+ "price_per_1000_decisions_usd": null,
1304
+ "schema_validity": 1.0,
1305
+ "schema_validity_strict": 1.0
1306
+ },
1307
+ "probability": {
1308
+ "accuracy": 0.4,
1309
+ "brier_mean": 0.7593166581141717,
1310
+ "calibration_n": 10,
1311
+ "cost_basis": [
1312
+ "local_gpu_no_provider_tariff"
1313
+ ],
1314
+ "coverage": 1.0,
1315
+ "ece": {
1316
+ "bins": [
1317
+ {
1318
+ "accuracy": null,
1319
+ "hi": 0.1,
1320
+ "lo": 0.0,
1321
+ "mean_confidence": null,
1322
+ "n": 0
1323
+ },
1324
+ {
1325
+ "accuracy": null,
1326
+ "hi": 0.2,
1327
+ "lo": 0.1,
1328
+ "mean_confidence": null,
1329
+ "n": 0
1330
+ },
1331
+ {
1332
+ "accuracy": null,
1333
+ "hi": 0.3,
1334
+ "lo": 0.2,
1335
+ "mean_confidence": null,
1336
+ "n": 0
1337
+ },
1338
+ {
1339
+ "accuracy": null,
1340
+ "hi": 0.4,
1341
+ "lo": 0.3,
1342
+ "mean_confidence": null,
1343
+ "n": 0
1344
+ },
1345
+ {
1346
+ "accuracy": 0.0,
1347
+ "hi": 0.5,
1348
+ "lo": 0.4,
1349
+ "mean_confidence": 0.4478049576282501,
1350
+ "n": 1
1351
+ },
1352
+ {
1353
+ "accuracy": 0.5,
1354
+ "hi": 0.6,
1355
+ "lo": 0.5,
1356
+ "mean_confidence": 0.5372944176197052,
1357
+ "n": 4
1358
+ },
1359
+ {
1360
+ "accuracy": 0.5,
1361
+ "hi": 0.7,
1362
+ "lo": 0.6,
1363
+ "mean_confidence": 0.6290208697319031,
1364
+ "n": 2
1365
+ },
1366
+ {
1367
+ "accuracy": null,
1368
+ "hi": 0.8,
1369
+ "lo": 0.7,
1370
+ "mean_confidence": null,
1371
+ "n": 0
1372
+ },
1373
+ {
1374
+ "accuracy": null,
1375
+ "hi": 0.9,
1376
+ "lo": 0.8,
1377
+ "mean_confidence": null,
1378
+ "n": 0
1379
+ },
1380
+ {
1381
+ "accuracy": 0.3333333333333333,
1382
+ "hi": 1.0,
1383
+ "lo": 0.9,
1384
+ "mean_confidence": 0.9370233019193014,
1385
+ "n": 3
1386
+ }
1387
+ ],
1388
+ "ece": 0.2666094273328781,
1389
+ "n": 10
1390
+ },
1391
+ "latency": {
1392
+ "n": 10,
1393
+ "p50_s": 0.11380846984684467,
1394
+ "p95_s": 0.12198328394442795
1395
+ },
1396
+ "latency_failures": {
1397
+ "n": 0,
1398
+ "p50_s": null,
1399
+ "p95_s": null
1400
+ },
1401
+ "n_attempted": 10,
1402
+ "n_correct": 4,
1403
+ "n_planned": 10,
1404
+ "n_renormalized": 0,
1405
+ "n_scorable": 10,
1406
+ "n_valid": 10,
1407
+ "operational_success": 1.0,
1408
+ "ordinal_mae": null,
1409
+ "price_per_1000_decisions_usd": null,
1410
+ "schema_validity": 1.0,
1411
+ "schema_validity_strict": 1.0
1412
+ },
1413
+ "routing_hard": {
1414
+ "accuracy": 1.0,
1415
+ "brier_mean": 0.0021777005726523737,
1416
+ "calibration_n": 5,
1417
+ "cost_basis": [
1418
+ "local_gpu_no_provider_tariff"
1419
+ ],
1420
+ "coverage": 1.0,
1421
+ "ece": {
1422
+ "bins": [
1423
+ {
1424
+ "accuracy": null,
1425
+ "hi": 0.1,
1426
+ "lo": 0.0,
1427
+ "mean_confidence": null,
1428
+ "n": 0
1429
+ },
1430
+ {
1431
+ "accuracy": null,
1432
+ "hi": 0.2,
1433
+ "lo": 0.1,
1434
+ "mean_confidence": null,
1435
+ "n": 0
1436
+ },
1437
+ {
1438
+ "accuracy": null,
1439
+ "hi": 0.3,
1440
+ "lo": 0.2,
1441
+ "mean_confidence": null,
1442
+ "n": 0
1443
+ },
1444
+ {
1445
+ "accuracy": null,
1446
+ "hi": 0.4,
1447
+ "lo": 0.3,
1448
+ "mean_confidence": null,
1449
+ "n": 0
1450
+ },
1451
+ {
1452
+ "accuracy": null,
1453
+ "hi": 0.5,
1454
+ "lo": 0.4,
1455
+ "mean_confidence": null,
1456
+ "n": 0
1457
+ },
1458
+ {
1459
+ "accuracy": null,
1460
+ "hi": 0.6,
1461
+ "lo": 0.5,
1462
+ "mean_confidence": null,
1463
+ "n": 0
1464
+ },
1465
+ {
1466
+ "accuracy": null,
1467
+ "hi": 0.7,
1468
+ "lo": 0.6,
1469
+ "mean_confidence": null,
1470
+ "n": 0
1471
+ },
1472
+ {
1473
+ "accuracy": null,
1474
+ "hi": 0.8,
1475
+ "lo": 0.7,
1476
+ "mean_confidence": null,
1477
+ "n": 0
1478
+ },
1479
+ {
1480
+ "accuracy": null,
1481
+ "hi": 0.9,
1482
+ "lo": 0.8,
1483
+ "mean_confidence": null,
1484
+ "n": 0
1485
+ },
1486
+ {
1487
+ "accuracy": 1.0,
1488
+ "hi": 1.0,
1489
+ "lo": 0.9,
1490
+ "mean_confidence": 0.9799626469612122,
1491
+ "n": 5
1492
+ }
1493
+ ],
1494
+ "ece": 0.020037353038787842,
1495
+ "n": 5
1496
+ },
1497
+ "latency": {
1498
+ "n": 5,
1499
+ "p50_s": 0.10272485390305519,
1500
+ "p95_s": 0.10452966764569283
1501
+ },
1502
+ "latency_failures": {
1503
+ "n": 0,
1504
+ "p50_s": null,
1505
+ "p95_s": null
1506
+ },
1507
+ "n_attempted": 5,
1508
+ "n_correct": 5,
1509
+ "n_planned": 5,
1510
+ "n_renormalized": 0,
1511
+ "n_scorable": 5,
1512
+ "n_valid": 5,
1513
+ "operational_success": 1.0,
1514
+ "ordinal_mae": null,
1515
+ "price_per_1000_decisions_usd": null,
1516
+ "schema_validity": 1.0,
1517
+ "schema_validity_strict": 1.0
1518
+ },
1519
+ "temporal_numeric": {
1520
+ "accuracy": 0.2,
1521
+ "brier_mean": 0.9778733222516895,
1522
+ "calibration_n": 15,
1523
+ "cost_basis": [
1524
+ "local_gpu_no_provider_tariff"
1525
+ ],
1526
+ "coverage": 1.0,
1527
+ "ece": {
1528
+ "bins": [
1529
+ {
1530
+ "accuracy": null,
1531
+ "hi": 0.1,
1532
+ "lo": 0.0,
1533
+ "mean_confidence": null,
1534
+ "n": 0
1535
+ },
1536
+ {
1537
+ "accuracy": null,
1538
+ "hi": 0.2,
1539
+ "lo": 0.1,
1540
+ "mean_confidence": null,
1541
+ "n": 0
1542
+ },
1543
+ {
1544
+ "accuracy": 0.0,
1545
+ "hi": 0.3,
1546
+ "lo": 0.2,
1547
+ "mean_confidence": 0.2788226008415222,
1548
+ "n": 2
1549
+ },
1550
+ {
1551
+ "accuracy": 0.0,
1552
+ "hi": 0.4,
1553
+ "lo": 0.3,
1554
+ "mean_confidence": 0.3616597553094228,
1555
+ "n": 3
1556
+ },
1557
+ {
1558
+ "accuracy": 0.5,
1559
+ "hi": 0.5,
1560
+ "lo": 0.4,
1561
+ "mean_confidence": 0.4627767503261566,
1562
+ "n": 2
1563
+ },
1564
+ {
1565
+ "accuracy": 0.0,
1566
+ "hi": 0.6,
1567
+ "lo": 0.5,
1568
+ "mean_confidence": 0.5329880118370056,
1569
+ "n": 3
1570
+ },
1571
+ {
1572
+ "accuracy": 1.0,
1573
+ "hi": 0.7,
1574
+ "lo": 0.6,
1575
+ "mean_confidence": 0.6510051488876343,
1576
+ "n": 2
1577
+ },
1578
+ {
1579
+ "accuracy": null,
1580
+ "hi": 0.8,
1581
+ "lo": 0.7,
1582
+ "mean_confidence": null,
1583
+ "n": 0
1584
+ },
1585
+ {
1586
+ "accuracy": 0.0,
1587
+ "hi": 0.9,
1588
+ "lo": 0.8,
1589
+ "mean_confidence": 0.887561023235321,
1590
+ "n": 1
1591
+ },
1592
+ {
1593
+ "accuracy": 0.0,
1594
+ "hi": 1.0,
1595
+ "lo": 0.9,
1596
+ "mean_confidence": 0.963701069355011,
1597
+ "n": 2
1598
+ }
1599
+ ],
1600
+ "ece": 0.45526585777600603,
1601
+ "n": 15
1602
+ },
1603
+ "latency": {
1604
+ "n": 15,
1605
+ "p50_s": 0.11358731612563133,
1606
+ "p95_s": 0.1201222252100706
1607
+ },
1608
+ "latency_failures": {
1609
+ "n": 0,
1610
+ "p50_s": null,
1611
+ "p95_s": null
1612
+ },
1613
+ "n_attempted": 15,
1614
+ "n_correct": 3,
1615
+ "n_planned": 15,
1616
+ "n_renormalized": 0,
1617
+ "n_scorable": 15,
1618
+ "n_valid": 15,
1619
+ "operational_success": 1.0,
1620
+ "ordinal_mae": 1.3456641659140587,
1621
+ "price_per_1000_decisions_usd": null,
1622
+ "schema_validity": 1.0,
1623
+ "schema_validity_strict": 1.0
1624
+ },
1625
+ "tradeoff": {
1626
+ "accuracy": 0.16666666666666666,
1627
+ "brier_mean": 1.1503974298690245,
1628
+ "calibration_n": 6,
1629
+ "cost_basis": [
1630
+ "local_gpu_no_provider_tariff"
1631
+ ],
1632
+ "coverage": 1.0,
1633
+ "ece": {
1634
+ "bins": [
1635
+ {
1636
+ "accuracy": null,
1637
+ "hi": 0.1,
1638
+ "lo": 0.0,
1639
+ "mean_confidence": null,
1640
+ "n": 0
1641
+ },
1642
+ {
1643
+ "accuracy": null,
1644
+ "hi": 0.2,
1645
+ "lo": 0.1,
1646
+ "mean_confidence": null,
1647
+ "n": 0
1648
+ },
1649
+ {
1650
+ "accuracy": null,
1651
+ "hi": 0.3,
1652
+ "lo": 0.2,
1653
+ "mean_confidence": null,
1654
+ "n": 0
1655
+ },
1656
+ {
1657
+ "accuracy": null,
1658
+ "hi": 0.4,
1659
+ "lo": 0.3,
1660
+ "mean_confidence": null,
1661
+ "n": 0
1662
+ },
1663
+ {
1664
+ "accuracy": null,
1665
+ "hi": 0.5,
1666
+ "lo": 0.4,
1667
+ "mean_confidence": null,
1668
+ "n": 0
1669
+ },
1670
+ {
1671
+ "accuracy": 0.0,
1672
+ "hi": 0.6,
1673
+ "lo": 0.5,
1674
+ "mean_confidence": 0.5204590559005737,
1675
+ "n": 1
1676
+ },
1677
+ {
1678
+ "accuracy": 0.3333333333333333,
1679
+ "hi": 0.7,
1680
+ "lo": 0.6,
1681
+ "mean_confidence": 0.6232326428095499,
1682
+ "n": 3
1683
+ },
1684
+ {
1685
+ "accuracy": null,
1686
+ "hi": 0.8,
1687
+ "lo": 0.7,
1688
+ "mean_confidence": null,
1689
+ "n": 0
1690
+ },
1691
+ {
1692
+ "accuracy": null,
1693
+ "hi": 0.9,
1694
+ "lo": 0.8,
1695
+ "mean_confidence": null,
1696
+ "n": 0
1697
+ },
1698
+ {
1699
+ "accuracy": 0.0,
1700
+ "hi": 1.0,
1701
+ "lo": 0.9,
1702
+ "mean_confidence": 0.9465669393539429,
1703
+ "n": 2
1704
+ }
1705
+ ],
1706
+ "ece": 0.5472151438395182,
1707
+ "n": 6
1708
+ },
1709
+ "latency": {
1710
+ "n": 6,
1711
+ "p50_s": 0.10983631946146488,
1712
+ "p95_s": 0.11511344648897648
1713
+ },
1714
+ "latency_failures": {
1715
+ "n": 0,
1716
+ "p50_s": null,
1717
+ "p95_s": null
1718
+ },
1719
+ "n_attempted": 6,
1720
+ "n_correct": 1,
1721
+ "n_planned": 6,
1722
+ "n_renormalized": 0,
1723
+ "n_scorable": 6,
1724
+ "n_valid": 6,
1725
+ "operational_success": 1.0,
1726
+ "ordinal_mae": null,
1727
+ "price_per_1000_decisions_usd": null,
1728
+ "schema_validity": 1.0,
1729
+ "schema_validity_strict": 1.0
1730
+ },
1731
+ "trap": {
1732
+ "accuracy": 0.625,
1733
+ "brier_mean": 0.3331839439462419,
1734
+ "calibration_n": 8,
1735
+ "cost_basis": [
1736
+ "local_gpu_no_provider_tariff"
1737
+ ],
1738
+ "coverage": 1.0,
1739
+ "ece": {
1740
+ "bins": [
1741
+ {
1742
+ "accuracy": null,
1743
+ "hi": 0.1,
1744
+ "lo": 0.0,
1745
+ "mean_confidence": null,
1746
+ "n": 0
1747
+ },
1748
+ {
1749
+ "accuracy": null,
1750
+ "hi": 0.2,
1751
+ "lo": 0.1,
1752
+ "mean_confidence": null,
1753
+ "n": 0
1754
+ },
1755
+ {
1756
+ "accuracy": null,
1757
+ "hi": 0.3,
1758
+ "lo": 0.2,
1759
+ "mean_confidence": null,
1760
+ "n": 0
1761
+ },
1762
+ {
1763
+ "accuracy": 0.0,
1764
+ "hi": 0.4,
1765
+ "lo": 0.3,
1766
+ "mean_confidence": 0.39877402782440186,
1767
+ "n": 1
1768
+ },
1769
+ {
1770
+ "accuracy": null,
1771
+ "hi": 0.5,
1772
+ "lo": 0.4,
1773
+ "mean_confidence": null,
1774
+ "n": 0
1775
+ },
1776
+ {
1777
+ "accuracy": 0.0,
1778
+ "hi": 0.6,
1779
+ "lo": 0.5,
1780
+ "mean_confidence": 0.5420189499855042,
1781
+ "n": 1
1782
+ },
1783
+ {
1784
+ "accuracy": 0.0,
1785
+ "hi": 0.7,
1786
+ "lo": 0.6,
1787
+ "mean_confidence": 0.6894591450691223,
1788
+ "n": 1
1789
+ },
1790
+ {
1791
+ "accuracy": null,
1792
+ "hi": 0.8,
1793
+ "lo": 0.7,
1794
+ "mean_confidence": null,
1795
+ "n": 0
1796
+ },
1797
+ {
1798
+ "accuracy": null,
1799
+ "hi": 0.9,
1800
+ "lo": 0.8,
1801
+ "mean_confidence": null,
1802
+ "n": 0
1803
+ },
1804
+ {
1805
+ "accuracy": 1.0,
1806
+ "hi": 1.0,
1807
+ "lo": 0.9,
1808
+ "mean_confidence": 0.9438392162322998,
1809
+ "n": 5
1810
+ }
1811
+ ],
1812
+ "ece": 0.23888200521469116,
1813
+ "n": 8
1814
+ },
1815
+ "latency": {
1816
+ "n": 8,
1817
+ "p50_s": 0.10395996645092964,
1818
+ "p95_s": 0.10869606416672467
1819
+ },
1820
+ "latency_failures": {
1821
+ "n": 0,
1822
+ "p50_s": null,
1823
+ "p95_s": null
1824
+ },
1825
+ "n_attempted": 8,
1826
+ "n_correct": 5,
1827
+ "n_planned": 8,
1828
+ "n_renormalized": 0,
1829
+ "n_scorable": 8,
1830
+ "n_valid": 8,
1831
+ "operational_success": 1.0,
1832
+ "ordinal_mae": 1.3499127477407455,
1833
+ "price_per_1000_decisions_usd": null,
1834
+ "schema_validity": 1.0,
1835
+ "schema_validity_strict": 1.0
1836
+ }
1837
+ },
1838
+ "price_per_1000_decisions_usd": null,
1839
+ "probability_sources": [
1840
+ "native"
1841
+ ],
1842
+ "schema_validity": 1.0,
1843
+ "schema_validity_strict": 1.0,
1844
+ "splits": {
1845
+ "public": {
1846
+ "accuracy": 0.42342342342342343,
1847
+ "brier_mean": 0.7731682475813886,
1848
+ "calibration_n": 111,
1849
+ "cost_basis": [
1850
+ "local_gpu_no_provider_tariff"
1851
+ ],
1852
+ "coverage": 1.0,
1853
+ "ece": {
1854
+ "bins": [
1855
+ {
1856
+ "accuracy": null,
1857
+ "hi": 0.1,
1858
+ "lo": 0.0,
1859
+ "mean_confidence": null,
1860
+ "n": 0
1861
+ },
1862
+ {
1863
+ "accuracy": null,
1864
+ "hi": 0.2,
1865
+ "lo": 0.1,
1866
+ "mean_confidence": null,
1867
+ "n": 0
1868
+ },
1869
+ {
1870
+ "accuracy": 0.0,
1871
+ "hi": 0.3,
1872
+ "lo": 0.2,
1873
+ "mean_confidence": 0.27849190682172775,
1874
+ "n": 4
1875
+ },
1876
+ {
1877
+ "accuracy": 0.125,
1878
+ "hi": 0.4,
1879
+ "lo": 0.3,
1880
+ "mean_confidence": 0.37296687439084053,
1881
+ "n": 8
1882
+ },
1883
+ {
1884
+ "accuracy": 0.4166666666666667,
1885
+ "hi": 0.5,
1886
+ "lo": 0.4,
1887
+ "mean_confidence": 0.4450460250178973,
1888
+ "n": 12
1889
+ },
1890
+ {
1891
+ "accuracy": 0.35,
1892
+ "hi": 0.6,
1893
+ "lo": 0.5,
1894
+ "mean_confidence": 0.5447314709424973,
1895
+ "n": 20
1896
+ },
1897
+ {
1898
+ "accuracy": 0.5,
1899
+ "hi": 0.7,
1900
+ "lo": 0.6,
1901
+ "mean_confidence": 0.6454204533781324,
1902
+ "n": 14
1903
+ },
1904
+ {
1905
+ "accuracy": 0.3333333333333333,
1906
+ "hi": 0.8,
1907
+ "lo": 0.7,
1908
+ "mean_confidence": 0.7522401312987009,
1909
+ "n": 6
1910
+ },
1911
+ {
1912
+ "accuracy": 0.25,
1913
+ "hi": 0.9,
1914
+ "lo": 0.8,
1915
+ "mean_confidence": 0.8441648408770561,
1916
+ "n": 8
1917
+ },
1918
+ {
1919
+ "accuracy": 0.5897435897435898,
1920
+ "hi": 1.0,
1921
+ "lo": 0.9,
1922
+ "mean_confidence": 0.965268883949671,
1923
+ "n": 39
1924
+ }
1925
+ ],
1926
+ "ece": 0.2818109610596219,
1927
+ "n": 111
1928
+ },
1929
+ "latency": {
1930
+ "n": 111,
1931
+ "p50_s": 0.11237071454524994,
1932
+ "p95_s": 0.14899961836636066
1933
+ },
1934
+ "latency_failures": {
1935
+ "n": 0,
1936
+ "p50_s": null,
1937
+ "p95_s": null
1938
+ },
1939
+ "n_attempted": 111,
1940
+ "n_correct": 47,
1941
+ "n_planned": 111,
1942
+ "n_renormalized": 0,
1943
+ "n_scorable": 111,
1944
+ "n_valid": 111,
1945
+ "operational_success": 1.0,
1946
+ "ordinal_mae": 0.7110752028723558,
1947
+ "price_per_1000_decisions_usd": null,
1948
+ "schema_validity": 1.0,
1949
+ "schema_validity_strict": 1.0
1950
+ }
1951
+ },
1952
+ "version": "jevbench-v1"
1953
+ },
1954
+ "original": {
1955
+ "accuracy": 0.8055555555555556,
1956
+ "brier_mean": 0.29652247410236326,
1957
+ "calibration_n": 72,
1958
+ "complete": true,
1959
+ "cost_basis": [
1960
+ "local_gpu_no_provider_tariff"
1961
+ ],
1962
+ "coverage": 1.0,
1963
+ "dataset_hash": "914276022fa5f0ea70e4231f2af98e68c67b274b61bea8c440b33890401cd2f9",
1964
+ "ece": {
1965
+ "bins": [
1966
+ {
1967
+ "accuracy": null,
1968
+ "hi": 0.1,
1969
+ "lo": 0.0,
1970
+ "mean_confidence": null,
1971
+ "n": 0
1972
+ },
1973
+ {
1974
+ "accuracy": null,
1975
+ "hi": 0.2,
1976
+ "lo": 0.1,
1977
+ "mean_confidence": null,
1978
+ "n": 0
1979
+ },
1980
+ {
1981
+ "accuracy": null,
1982
+ "hi": 0.3,
1983
+ "lo": 0.2,
1984
+ "mean_confidence": null,
1985
+ "n": 0
1986
+ },
1987
+ {
1988
+ "accuracy": 0.5,
1989
+ "hi": 0.4,
1990
+ "lo": 0.3,
1991
+ "mean_confidence": 0.3444499373435974,
1992
+ "n": 2
1993
+ },
1994
+ {
1995
+ "accuracy": 1.0,
1996
+ "hi": 0.5,
1997
+ "lo": 0.4,
1998
+ "mean_confidence": 0.4325794279575348,
1999
+ "n": 1
2000
+ },
2001
+ {
2002
+ "accuracy": 0.42857142857142855,
2003
+ "hi": 0.6,
2004
+ "lo": 0.5,
2005
+ "mean_confidence": 0.5371253064700535,
2006
+ "n": 7
2007
+ },
2008
+ {
2009
+ "accuracy": 0.6666666666666666,
2010
+ "hi": 0.7,
2011
+ "lo": 0.6,
2012
+ "mean_confidence": 0.6374844577577379,
2013
+ "n": 9
2014
+ },
2015
+ {
2016
+ "accuracy": 0.8571428571428571,
2017
+ "hi": 0.8,
2018
+ "lo": 0.7,
2019
+ "mean_confidence": 0.7471651860645839,
2020
+ "n": 7
2021
+ },
2022
+ {
2023
+ "accuracy": 0.9,
2024
+ "hi": 0.9,
2025
+ "lo": 0.8,
2026
+ "mean_confidence": 0.8518721401691437,
2027
+ "n": 10
2028
+ },
2029
+ {
2030
+ "accuracy": 0.8888888888888888,
2031
+ "hi": 1.0,
2032
+ "lo": 0.9,
2033
+ "mean_confidence": 0.9730205883582433,
2034
+ "n": 36
2035
+ }
2036
+ ],
2037
+ "ece": 0.08584584999415612,
2038
+ "n": 72
2039
+ },
2040
+ "latency": {
2041
+ "n": 72,
2042
+ "p50_s": 0.07467062026262283,
2043
+ "p95_s": 0.1292133398354054
2044
+ },
2045
+ "latency_failures": {
2046
+ "n": 0,
2047
+ "p50_s": null,
2048
+ "p95_s": null
2049
+ },
2050
+ "ledger_charged_usd": null,
2051
+ "macro_accuracy": 0.8055555555555555,
2052
+ "model_identities": [
2053
+ "pcdm:native:runs/tm1b"
2054
+ ],
2055
+ "n_attempted": 72,
2056
+ "n_correct": 58,
2057
+ "n_planned": 72,
2058
+ "n_renormalized": 0,
2059
+ "n_scorable": 72,
2060
+ "n_valid": 72,
2061
+ "operational_success": 1.0,
2062
+ "ordinal_mae": 0.39357847470091656,
2063
+ "paraphrase_consistency": {
2064
+ "agree": 30,
2065
+ "agreement": 0.8333333333333334,
2066
+ "both_correct_rate_all_pairs": 0.7222222222222222,
2067
+ "both_valid": 36,
2068
+ "pairs": 36
2069
+ },
2070
+ "per_family": {
2071
+ "adequacy": {
2072
+ "accuracy": 0.6666666666666666,
2073
+ "brier_mean": 0.40836962254148307,
2074
+ "calibration_n": 12,
2075
+ "cost_basis": [
2076
+ "local_gpu_no_provider_tariff"
2077
+ ],
2078
+ "coverage": 1.0,
2079
+ "ece": {
2080
+ "bins": [
2081
+ {
2082
+ "accuracy": null,
2083
+ "hi": 0.1,
2084
+ "lo": 0.0,
2085
+ "mean_confidence": null,
2086
+ "n": 0
2087
+ },
2088
+ {
2089
+ "accuracy": null,
2090
+ "hi": 0.2,
2091
+ "lo": 0.1,
2092
+ "mean_confidence": null,
2093
+ "n": 0
2094
+ },
2095
+ {
2096
+ "accuracy": null,
2097
+ "hi": 0.3,
2098
+ "lo": 0.2,
2099
+ "mean_confidence": null,
2100
+ "n": 0
2101
+ },
2102
+ {
2103
+ "accuracy": null,
2104
+ "hi": 0.4,
2105
+ "lo": 0.3,
2106
+ "mean_confidence": null,
2107
+ "n": 0
2108
+ },
2109
+ {
2110
+ "accuracy": null,
2111
+ "hi": 0.5,
2112
+ "lo": 0.4,
2113
+ "mean_confidence": null,
2114
+ "n": 0
2115
+ },
2116
+ {
2117
+ "accuracy": 0.0,
2118
+ "hi": 0.6,
2119
+ "lo": 0.5,
2120
+ "mean_confidence": 0.5817238092422485,
2121
+ "n": 1
2122
+ },
2123
+ {
2124
+ "accuracy": 0.5,
2125
+ "hi": 0.7,
2126
+ "lo": 0.6,
2127
+ "mean_confidence": 0.6328441649675369,
2128
+ "n": 4
2129
+ },
2130
+ {
2131
+ "accuracy": null,
2132
+ "hi": 0.8,
2133
+ "lo": 0.7,
2134
+ "mean_confidence": null,
2135
+ "n": 0
2136
+ },
2137
+ {
2138
+ "accuracy": 1.0,
2139
+ "hi": 0.9,
2140
+ "lo": 0.8,
2141
+ "mean_confidence": 0.8540955781936646,
2142
+ "n": 1
2143
+ },
2144
+ {
2145
+ "accuracy": 0.8333333333333334,
2146
+ "hi": 1.0,
2147
+ "lo": 0.9,
2148
+ "mean_confidence": 0.972666323184967,
2149
+ "n": 6
2150
+ }
2151
+ ],
2152
+ "ece": 0.1745835691690445,
2153
+ "n": 12
2154
+ },
2155
+ "latency": {
2156
+ "n": 12,
2157
+ "p50_s": 0.07050643116235733,
2158
+ "p95_s": 0.08810084797441958
2159
+ },
2160
+ "latency_failures": {
2161
+ "n": 0,
2162
+ "p50_s": null,
2163
+ "p95_s": null
2164
+ },
2165
+ "n_attempted": 12,
2166
+ "n_correct": 8,
2167
+ "n_planned": 12,
2168
+ "n_renormalized": 0,
2169
+ "n_scorable": 12,
2170
+ "n_valid": 12,
2171
+ "operational_success": 1.0,
2172
+ "ordinal_mae": null,
2173
+ "price_per_1000_decisions_usd": null,
2174
+ "schema_validity": 1.0,
2175
+ "schema_validity_strict": 1.0
2176
+ },
2177
+ "extraction": {
2178
+ "accuracy": 1.0,
2179
+ "brier_mean": 0.02007323946682944,
2180
+ "calibration_n": 12,
2181
+ "cost_basis": [
2182
+ "local_gpu_no_provider_tariff"
2183
+ ],
2184
+ "coverage": 1.0,
2185
+ "ece": {
2186
+ "bins": [
2187
+ {
2188
+ "accuracy": null,
2189
+ "hi": 0.1,
2190
+ "lo": 0.0,
2191
+ "mean_confidence": null,
2192
+ "n": 0
2193
+ },
2194
+ {
2195
+ "accuracy": null,
2196
+ "hi": 0.2,
2197
+ "lo": 0.1,
2198
+ "mean_confidence": null,
2199
+ "n": 0
2200
+ },
2201
+ {
2202
+ "accuracy": null,
2203
+ "hi": 0.3,
2204
+ "lo": 0.2,
2205
+ "mean_confidence": null,
2206
+ "n": 0
2207
+ },
2208
+ {
2209
+ "accuracy": null,
2210
+ "hi": 0.4,
2211
+ "lo": 0.3,
2212
+ "mean_confidence": null,
2213
+ "n": 0
2214
+ },
2215
+ {
2216
+ "accuracy": null,
2217
+ "hi": 0.5,
2218
+ "lo": 0.4,
2219
+ "mean_confidence": null,
2220
+ "n": 0
2221
+ },
2222
+ {
2223
+ "accuracy": null,
2224
+ "hi": 0.6,
2225
+ "lo": 0.5,
2226
+ "mean_confidence": null,
2227
+ "n": 0
2228
+ },
2229
+ {
2230
+ "accuracy": null,
2231
+ "hi": 0.7,
2232
+ "lo": 0.6,
2233
+ "mean_confidence": null,
2234
+ "n": 0
2235
+ },
2236
+ {
2237
+ "accuracy": 1.0,
2238
+ "hi": 0.8,
2239
+ "lo": 0.7,
2240
+ "mean_confidence": 0.7408376932144165,
2241
+ "n": 1
2242
+ },
2243
+ {
2244
+ "accuracy": 1.0,
2245
+ "hi": 0.9,
2246
+ "lo": 0.8,
2247
+ "mean_confidence": 0.8441984951496124,
2248
+ "n": 2
2249
+ },
2250
+ {
2251
+ "accuracy": 1.0,
2252
+ "hi": 1.0,
2253
+ "lo": 0.9,
2254
+ "mean_confidence": 0.968041201432546,
2255
+ "n": 9
2256
+ }
2257
+ ],
2258
+ "ece": 0.07153287529945371,
2259
+ "n": 12
2260
+ },
2261
+ "latency": {
2262
+ "n": 12,
2263
+ "p50_s": 0.05768968164920807,
2264
+ "p95_s": 0.08274002466350794
2265
+ },
2266
+ "latency_failures": {
2267
+ "n": 0,
2268
+ "p50_s": null,
2269
+ "p95_s": null
2270
+ },
2271
+ "n_attempted": 12,
2272
+ "n_correct": 12,
2273
+ "n_planned": 12,
2274
+ "n_renormalized": 0,
2275
+ "n_scorable": 12,
2276
+ "n_valid": 12,
2277
+ "operational_success": 1.0,
2278
+ "ordinal_mae": null,
2279
+ "price_per_1000_decisions_usd": null,
2280
+ "schema_validity": 1.0,
2281
+ "schema_validity_strict": 1.0
2282
+ },
2283
+ "intent": {
2284
+ "accuracy": 0.9166666666666666,
2285
+ "brier_mean": 0.22698352684655446,
2286
+ "calibration_n": 12,
2287
+ "cost_basis": [
2288
+ "local_gpu_no_provider_tariff"
2289
+ ],
2290
+ "coverage": 1.0,
2291
+ "ece": {
2292
+ "bins": [
2293
+ {
2294
+ "accuracy": null,
2295
+ "hi": 0.1,
2296
+ "lo": 0.0,
2297
+ "mean_confidence": null,
2298
+ "n": 0
2299
+ },
2300
+ {
2301
+ "accuracy": null,
2302
+ "hi": 0.2,
2303
+ "lo": 0.1,
2304
+ "mean_confidence": null,
2305
+ "n": 0
2306
+ },
2307
+ {
2308
+ "accuracy": null,
2309
+ "hi": 0.3,
2310
+ "lo": 0.2,
2311
+ "mean_confidence": null,
2312
+ "n": 0
2313
+ },
2314
+ {
2315
+ "accuracy": null,
2316
+ "hi": 0.4,
2317
+ "lo": 0.3,
2318
+ "mean_confidence": null,
2319
+ "n": 0
2320
+ },
2321
+ {
2322
+ "accuracy": null,
2323
+ "hi": 0.5,
2324
+ "lo": 0.4,
2325
+ "mean_confidence": null,
2326
+ "n": 0
2327
+ },
2328
+ {
2329
+ "accuracy": 1.0,
2330
+ "hi": 0.6,
2331
+ "lo": 0.5,
2332
+ "mean_confidence": 0.5113146007061005,
2333
+ "n": 2
2334
+ },
2335
+ {
2336
+ "accuracy": null,
2337
+ "hi": 0.7,
2338
+ "lo": 0.6,
2339
+ "mean_confidence": null,
2340
+ "n": 0
2341
+ },
2342
+ {
2343
+ "accuracy": 1.0,
2344
+ "hi": 0.8,
2345
+ "lo": 0.7,
2346
+ "mean_confidence": 0.7275521159172058,
2347
+ "n": 1
2348
+ },
2349
+ {
2350
+ "accuracy": 1.0,
2351
+ "hi": 0.9,
2352
+ "lo": 0.8,
2353
+ "mean_confidence": 0.8921765089035034,
2354
+ "n": 1
2355
+ },
2356
+ {
2357
+ "accuracy": 0.875,
2358
+ "hi": 1.0,
2359
+ "lo": 0.9,
2360
+ "mean_confidence": 0.9776975587010384,
2361
+ "n": 8
2362
+ }
2363
+ ],
2364
+ "ece": 0.1816018869479497,
2365
+ "n": 12
2366
+ },
2367
+ "latency": {
2368
+ "n": 12,
2369
+ "p50_s": 0.11113478429615498,
2370
+ "p95_s": 0.19621948227286334
2371
+ },
2372
+ "latency_failures": {
2373
+ "n": 0,
2374
+ "p50_s": null,
2375
+ "p95_s": null
2376
+ },
2377
+ "n_attempted": 12,
2378
+ "n_correct": 11,
2379
+ "n_planned": 12,
2380
+ "n_renormalized": 0,
2381
+ "n_scorable": 12,
2382
+ "n_valid": 12,
2383
+ "operational_success": 1.0,
2384
+ "ordinal_mae": null,
2385
+ "price_per_1000_decisions_usd": null,
2386
+ "schema_validity": 1.0,
2387
+ "schema_validity_strict": 1.0
2388
+ },
2389
+ "ordinal": {
2390
+ "accuracy": 0.6666666666666666,
2391
+ "brier_mean": 0.387267867931643,
2392
+ "calibration_n": 12,
2393
+ "cost_basis": [
2394
+ "local_gpu_no_provider_tariff"
2395
+ ],
2396
+ "coverage": 1.0,
2397
+ "ece": {
2398
+ "bins": [
2399
+ {
2400
+ "accuracy": null,
2401
+ "hi": 0.1,
2402
+ "lo": 0.0,
2403
+ "mean_confidence": null,
2404
+ "n": 0
2405
+ },
2406
+ {
2407
+ "accuracy": null,
2408
+ "hi": 0.2,
2409
+ "lo": 0.1,
2410
+ "mean_confidence": null,
2411
+ "n": 0
2412
+ },
2413
+ {
2414
+ "accuracy": null,
2415
+ "hi": 0.3,
2416
+ "lo": 0.2,
2417
+ "mean_confidence": null,
2418
+ "n": 0
2419
+ },
2420
+ {
2421
+ "accuracy": 0.5,
2422
+ "hi": 0.4,
2423
+ "lo": 0.3,
2424
+ "mean_confidence": 0.3444499373435974,
2425
+ "n": 2
2426
+ },
2427
+ {
2428
+ "accuracy": 1.0,
2429
+ "hi": 0.5,
2430
+ "lo": 0.4,
2431
+ "mean_confidence": 0.4325794279575348,
2432
+ "n": 1
2433
+ },
2434
+ {
2435
+ "accuracy": 0.0,
2436
+ "hi": 0.6,
2437
+ "lo": 0.5,
2438
+ "mean_confidence": 0.5386167963345846,
2439
+ "n": 3
2440
+ },
2441
+ {
2442
+ "accuracy": 1.0,
2443
+ "hi": 0.7,
2444
+ "lo": 0.6,
2445
+ "mean_confidence": 0.6215285360813141,
2446
+ "n": 2
2447
+ },
2448
+ {
2449
+ "accuracy": null,
2450
+ "hi": 0.8,
2451
+ "lo": 0.7,
2452
+ "mean_confidence": null,
2453
+ "n": 0
2454
+ },
2455
+ {
2456
+ "accuracy": 1.0,
2457
+ "hi": 0.9,
2458
+ "lo": 0.8,
2459
+ "mean_confidence": 0.839689165353775,
2460
+ "n": 4
2461
+ },
2462
+ {
2463
+ "accuracy": null,
2464
+ "hi": 1.0,
2465
+ "lo": 0.9,
2466
+ "mean_confidence": null,
2467
+ "n": 0
2468
+ }
2469
+ ],
2470
+ "ece": 0.32437977939844137,
2471
+ "n": 12
2472
+ },
2473
+ "latency": {
2474
+ "n": 12,
2475
+ "p50_s": 0.07719611376523972,
2476
+ "p95_s": 0.11410039383918047
2477
+ },
2478
+ "latency_failures": {
2479
+ "n": 0,
2480
+ "p50_s": null,
2481
+ "p95_s": null
2482
+ },
2483
+ "n_attempted": 12,
2484
+ "n_correct": 8,
2485
+ "n_planned": 12,
2486
+ "n_renormalized": 0,
2487
+ "n_scorable": 12,
2488
+ "n_valid": 12,
2489
+ "operational_success": 1.0,
2490
+ "ordinal_mae": 0.39357847470091656,
2491
+ "price_per_1000_decisions_usd": null,
2492
+ "schema_validity": 1.0,
2493
+ "schema_validity_strict": 1.0
2494
+ },
2495
+ "policy": {
2496
+ "accuracy": 0.75,
2497
+ "brier_mean": 0.3966635065443795,
2498
+ "calibration_n": 12,
2499
+ "cost_basis": [
2500
+ "local_gpu_no_provider_tariff"
2501
+ ],
2502
+ "coverage": 1.0,
2503
+ "ece": {
2504
+ "bins": [
2505
+ {
2506
+ "accuracy": null,
2507
+ "hi": 0.1,
2508
+ "lo": 0.0,
2509
+ "mean_confidence": null,
2510
+ "n": 0
2511
+ },
2512
+ {
2513
+ "accuracy": null,
2514
+ "hi": 0.2,
2515
+ "lo": 0.1,
2516
+ "mean_confidence": null,
2517
+ "n": 0
2518
+ },
2519
+ {
2520
+ "accuracy": null,
2521
+ "hi": 0.3,
2522
+ "lo": 0.2,
2523
+ "mean_confidence": null,
2524
+ "n": 0
2525
+ },
2526
+ {
2527
+ "accuracy": null,
2528
+ "hi": 0.4,
2529
+ "lo": 0.3,
2530
+ "mean_confidence": null,
2531
+ "n": 0
2532
+ },
2533
+ {
2534
+ "accuracy": null,
2535
+ "hi": 0.5,
2536
+ "lo": 0.4,
2537
+ "mean_confidence": null,
2538
+ "n": 0
2539
+ },
2540
+ {
2541
+ "accuracy": null,
2542
+ "hi": 0.6,
2543
+ "lo": 0.5,
2544
+ "mean_confidence": null,
2545
+ "n": 0
2546
+ },
2547
+ {
2548
+ "accuracy": 0.5,
2549
+ "hi": 0.7,
2550
+ "lo": 0.6,
2551
+ "mean_confidence": 0.6735906898975372,
2552
+ "n": 2
2553
+ },
2554
+ {
2555
+ "accuracy": null,
2556
+ "hi": 0.8,
2557
+ "lo": 0.7,
2558
+ "mean_confidence": null,
2559
+ "n": 0
2560
+ },
2561
+ {
2562
+ "accuracy": 1.0,
2563
+ "hi": 0.9,
2564
+ "lo": 0.8,
2565
+ "mean_confidence": 0.859863817691803,
2566
+ "n": 1
2567
+ },
2568
+ {
2569
+ "accuracy": 0.7777777777777778,
2570
+ "hi": 1.0,
2571
+ "lo": 0.9,
2572
+ "mean_confidence": 0.9743029342757331,
2573
+ "n": 9
2574
+ }
2575
+ ],
2576
+ "ece": 0.18800366421540576,
2577
+ "n": 12
2578
+ },
2579
+ "latency": {
2580
+ "n": 12,
2581
+ "p50_s": 0.10365468636155128,
2582
+ "p95_s": 0.15637512318789956
2583
+ },
2584
+ "latency_failures": {
2585
+ "n": 0,
2586
+ "p50_s": null,
2587
+ "p95_s": null
2588
+ },
2589
+ "n_attempted": 12,
2590
+ "n_correct": 9,
2591
+ "n_planned": 12,
2592
+ "n_renormalized": 0,
2593
+ "n_scorable": 12,
2594
+ "n_valid": 12,
2595
+ "operational_success": 1.0,
2596
+ "ordinal_mae": null,
2597
+ "price_per_1000_decisions_usd": null,
2598
+ "schema_validity": 1.0,
2599
+ "schema_validity_strict": 1.0
2600
+ },
2601
+ "routing": {
2602
+ "accuracy": 0.8333333333333334,
2603
+ "brier_mean": 0.3397770812832901,
2604
+ "calibration_n": 12,
2605
+ "cost_basis": [
2606
+ "local_gpu_no_provider_tariff"
2607
+ ],
2608
+ "coverage": 1.0,
2609
+ "ece": {
2610
+ "bins": [
2611
+ {
2612
+ "accuracy": null,
2613
+ "hi": 0.1,
2614
+ "lo": 0.0,
2615
+ "mean_confidence": null,
2616
+ "n": 0
2617
+ },
2618
+ {
2619
+ "accuracy": null,
2620
+ "hi": 0.2,
2621
+ "lo": 0.1,
2622
+ "mean_confidence": null,
2623
+ "n": 0
2624
+ },
2625
+ {
2626
+ "accuracy": null,
2627
+ "hi": 0.3,
2628
+ "lo": 0.2,
2629
+ "mean_confidence": null,
2630
+ "n": 0
2631
+ },
2632
+ {
2633
+ "accuracy": null,
2634
+ "hi": 0.4,
2635
+ "lo": 0.3,
2636
+ "mean_confidence": null,
2637
+ "n": 0
2638
+ },
2639
+ {
2640
+ "accuracy": null,
2641
+ "hi": 0.5,
2642
+ "lo": 0.4,
2643
+ "mean_confidence": null,
2644
+ "n": 0
2645
+ },
2646
+ {
2647
+ "accuracy": 1.0,
2648
+ "hi": 0.6,
2649
+ "lo": 0.5,
2650
+ "mean_confidence": 0.5396737456321716,
2651
+ "n": 1
2652
+ },
2653
+ {
2654
+ "accuracy": 1.0,
2655
+ "hi": 0.7,
2656
+ "lo": 0.6,
2657
+ "mean_confidence": 0.6157450079917908,
2658
+ "n": 1
2659
+ },
2660
+ {
2661
+ "accuracy": 0.8,
2662
+ "hi": 0.8,
2663
+ "lo": 0.7,
2664
+ "mean_confidence": 0.7523532986640931,
2665
+ "n": 5
2666
+ },
2667
+ {
2668
+ "accuracy": 0.0,
2669
+ "hi": 0.9,
2670
+ "lo": 0.8,
2671
+ "mean_confidence": 0.8654318451881409,
2672
+ "n": 1
2673
+ },
2674
+ {
2675
+ "accuracy": 1.0,
2676
+ "hi": 1.0,
2677
+ "lo": 0.9,
2678
+ "mean_confidence": 0.9725163877010345,
2679
+ "n": 4
2680
+ }
2681
+ ],
2682
+ "ece": 0.17151508728663126,
2683
+ "n": 12
2684
+ },
2685
+ "latency": {
2686
+ "n": 12,
2687
+ "p50_s": 0.053897660225629807,
2688
+ "p95_s": 0.08854644279927014
2689
+ },
2690
+ "latency_failures": {
2691
+ "n": 0,
2692
+ "p50_s": null,
2693
+ "p95_s": null
2694
+ },
2695
+ "n_attempted": 12,
2696
+ "n_correct": 10,
2697
+ "n_planned": 12,
2698
+ "n_renormalized": 0,
2699
+ "n_scorable": 12,
2700
+ "n_valid": 12,
2701
+ "operational_success": 1.0,
2702
+ "ordinal_mae": null,
2703
+ "price_per_1000_decisions_usd": null,
2704
+ "schema_validity": 1.0,
2705
+ "schema_validity_strict": 1.0
2706
+ }
2707
+ },
2708
+ "price_per_1000_decisions_usd": null,
2709
+ "probability_sources": [
2710
+ "native"
2711
+ ],
2712
+ "schema_validity": 1.0,
2713
+ "schema_validity_strict": 1.0,
2714
+ "splits": {
2715
+ "public": {
2716
+ "accuracy": 0.8055555555555556,
2717
+ "brier_mean": 0.29652247410236326,
2718
+ "calibration_n": 72,
2719
+ "cost_basis": [
2720
+ "local_gpu_no_provider_tariff"
2721
+ ],
2722
+ "coverage": 1.0,
2723
+ "ece": {
2724
+ "bins": [
2725
+ {
2726
+ "accuracy": null,
2727
+ "hi": 0.1,
2728
+ "lo": 0.0,
2729
+ "mean_confidence": null,
2730
+ "n": 0
2731
+ },
2732
+ {
2733
+ "accuracy": null,
2734
+ "hi": 0.2,
2735
+ "lo": 0.1,
2736
+ "mean_confidence": null,
2737
+ "n": 0
2738
+ },
2739
+ {
2740
+ "accuracy": null,
2741
+ "hi": 0.3,
2742
+ "lo": 0.2,
2743
+ "mean_confidence": null,
2744
+ "n": 0
2745
+ },
2746
+ {
2747
+ "accuracy": 0.5,
2748
+ "hi": 0.4,
2749
+ "lo": 0.3,
2750
+ "mean_confidence": 0.3444499373435974,
2751
+ "n": 2
2752
+ },
2753
+ {
2754
+ "accuracy": 1.0,
2755
+ "hi": 0.5,
2756
+ "lo": 0.4,
2757
+ "mean_confidence": 0.4325794279575348,
2758
+ "n": 1
2759
+ },
2760
+ {
2761
+ "accuracy": 0.42857142857142855,
2762
+ "hi": 0.6,
2763
+ "lo": 0.5,
2764
+ "mean_confidence": 0.5371253064700535,
2765
+ "n": 7
2766
+ },
2767
+ {
2768
+ "accuracy": 0.6666666666666666,
2769
+ "hi": 0.7,
2770
+ "lo": 0.6,
2771
+ "mean_confidence": 0.6374844577577379,
2772
+ "n": 9
2773
+ },
2774
+ {
2775
+ "accuracy": 0.8571428571428571,
2776
+ "hi": 0.8,
2777
+ "lo": 0.7,
2778
+ "mean_confidence": 0.7471651860645839,
2779
+ "n": 7
2780
+ },
2781
+ {
2782
+ "accuracy": 0.9,
2783
+ "hi": 0.9,
2784
+ "lo": 0.8,
2785
+ "mean_confidence": 0.8518721401691437,
2786
+ "n": 10
2787
+ },
2788
+ {
2789
+ "accuracy": 0.8888888888888888,
2790
+ "hi": 1.0,
2791
+ "lo": 0.9,
2792
+ "mean_confidence": 0.9730205883582433,
2793
+ "n": 36
2794
+ }
2795
+ ],
2796
+ "ece": 0.08584584999415612,
2797
+ "n": 72
2798
+ },
2799
+ "latency": {
2800
+ "n": 72,
2801
+ "p50_s": 0.07467062026262283,
2802
+ "p95_s": 0.1292133398354054
2803
+ },
2804
+ "latency_failures": {
2805
+ "n": 0,
2806
+ "p50_s": null,
2807
+ "p95_s": null
2808
+ },
2809
+ "n_attempted": 72,
2810
+ "n_correct": 58,
2811
+ "n_planned": 72,
2812
+ "n_renormalized": 0,
2813
+ "n_scorable": 72,
2814
+ "n_valid": 72,
2815
+ "operational_success": 1.0,
2816
+ "ordinal_mae": 0.39357847470091656,
2817
+ "price_per_1000_decisions_usd": null,
2818
+ "schema_validity": 1.0,
2819
+ "schema_validity_strict": 1.0
2820
+ }
2821
+ },
2822
+ "version": "jevbench-v1"
2823
+ }
2824
+ }
probe_results.json ADDED
@@ -0,0 +1,155 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "T": 1.1242100350620867,
3
+ "null_offset": 0.0,
4
+ "calib_sets": [
5
+ "val"
6
+ ],
7
+ "val_nll": 0.33989632085942884,
8
+ "args": {
9
+ "name": "probe_tm1b",
10
+ "backbone": "Qwen/Qwen3-4B-Base",
11
+ "lora_layers": 8,
12
+ "lora_r": 16,
13
+ "steps": 12000,
14
+ "bs": 64,
15
+ "grad_accum": 1,
16
+ "lr": 0.0003,
17
+ "lora_lr": 0.0001,
18
+ "seed": 0,
19
+ "no_hybrid": false,
20
+ "no_cand_null": false,
21
+ "no_null": false,
22
+ "null": "factored",
23
+ "hard_only": false,
24
+ "mix": "full",
25
+ "eval_every": 4000,
26
+ "val_every": 1000,
27
+ "ckpt_every": 2000,
28
+ "eval_bs": 16,
29
+ "device": "auto",
30
+ "wandb": false,
31
+ "hf_repo": null,
32
+ "ckpt_upload": false,
33
+ "data": "data_probe",
34
+ "extra_data": null,
35
+ "eval_cap": null,
36
+ "calib_sets": null,
37
+ "family_weights": null,
38
+ "bucket_map": null,
39
+ "null_aug": null,
40
+ "smoke": false,
41
+ "eval_limit": 0,
42
+ "tap_layer": 26,
43
+ "zscore": false,
44
+ "eval_only": true,
45
+ "dump_logits": null,
46
+ "tower_d": 512,
47
+ "tower_layers": 2,
48
+ "tower_heads": 8,
49
+ "joint": false,
50
+ "listwise": false,
51
+ "extra_tap": 0,
52
+ "init_from": null,
53
+ "readout": "native",
54
+ "nc_head": "n3",
55
+ "nc_render": "letters_nonull",
56
+ "score_head": "choice",
57
+ "noul_head": "bern",
58
+ "qtype_filter": null,
59
+ "ordinal_smooth": 0.0,
60
+ "max_state": 1024,
61
+ "perm_lambda": 0.0,
62
+ "no_shuffle": false,
63
+ "zero_shot": false,
64
+ "shots": 0,
65
+ "cand_encoder": "backbone",
66
+ "head": "mlp",
67
+ "z_dim": 128,
68
+ "z_probes": 8,
69
+ "tiny_layers": 2,
70
+ "distill_alpha": 1.0,
71
+ "distill_beta": 0.0,
72
+ "distill_T": 2.0,
73
+ "delta_gamma": 0.0
74
+ },
75
+ "eval": {
76
+ "mmlu_choicesonly": {
77
+ "raw": {
78
+ "nll": 2.039828592927867,
79
+ "brier": 0.8168081995898367,
80
+ "jsd": 0.4534946954898565,
81
+ "acc": 0.29,
82
+ "acc_k": 0.29083333333333333,
83
+ "ece": 0.04304502696419755,
84
+ "sel_acc80": 0.3229166666666667,
85
+ "coverage": 0.9966666666666667,
86
+ "false_abstain": 0.0033333333333333335,
87
+ "null_recall": "n/a"
88
+ },
89
+ "scaled": {
90
+ "nll": 2.0257116311704886,
91
+ "brier": 0.8150993429201715,
92
+ "jsd": 0.4567105766310495,
93
+ "acc": 0.29,
94
+ "acc_k": 0.29083333333333333,
95
+ "ece": 0.03703421314557393,
96
+ "sel_acc80": 0.32395833333333335,
97
+ "coverage": 0.9966666666666667,
98
+ "false_abstain": 0.0033333333333333335,
99
+ "null_recall": "n/a"
100
+ }
101
+ },
102
+ "mmlu_pro": {
103
+ "raw": {
104
+ "nll": 1.6070274236514384,
105
+ "brier": 0.6767720553604334,
106
+ "jsd": 0.3678978835677458,
107
+ "acc": 0.455,
108
+ "acc_k": 0.4583333333333333,
109
+ "ece": 0.04268285360187291,
110
+ "sel_acc80": 0.5166666666666667,
111
+ "coverage": 0.9958333333333333,
112
+ "false_abstain": 0.004166666666666667,
113
+ "null_recall": "n/a"
114
+ },
115
+ "scaled": {
116
+ "nll": 1.6083517592633751,
117
+ "brier": 0.6793089765777306,
118
+ "jsd": 0.3746115181488698,
119
+ "acc": 0.455,
120
+ "acc_k": 0.4583333333333333,
121
+ "ece": 0.05830767397458354,
122
+ "sel_acc80": 0.5166666666666667,
123
+ "coverage": 0.9958333333333333,
124
+ "false_abstain": 0.004166666666666667,
125
+ "null_recall": "n/a"
126
+ }
127
+ },
128
+ "mmlu_shuffledq": {
129
+ "raw": {
130
+ "nll": 2.0994824466839295,
131
+ "brier": 0.839300334664014,
132
+ "jsd": 0.4620919566125671,
133
+ "acc": 0.2633333333333333,
134
+ "acc_k": 0.265,
135
+ "ece": 0.06165247778718671,
136
+ "sel_acc80": 0.29270833333333335,
137
+ "coverage": 0.9933333333333333,
138
+ "false_abstain": 0.006666666666666667,
139
+ "null_recall": "n/a"
140
+ },
141
+ "scaled": {
142
+ "nll": 2.0802900159544735,
143
+ "brier": 0.8357595643299349,
144
+ "jsd": 0.46472209676556786,
145
+ "acc": 0.2633333333333333,
146
+ "acc_k": 0.265,
147
+ "ece": 0.03918854349603254,
148
+ "sel_acc80": 0.29270833333333335,
149
+ "coverage": 0.9933333333333333,
150
+ "false_abstain": 0.006666666666666667,
151
+ "null_recall": "n/a"
152
+ }
153
+ }
154
+ }
155
+ }
results.json ADDED
@@ -0,0 +1,2248 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "T": 1.1242100350620867,
3
+ "null_offset": 0.0,
4
+ "calib_sets": [
5
+ "val"
6
+ ],
7
+ "val_nll": 0.33989632085942884,
8
+ "args": {
9
+ "name": "tm1b",
10
+ "backbone": "Qwen/Qwen3-4B-Base",
11
+ "lora_layers": 8,
12
+ "lora_r": 16,
13
+ "steps": 12000,
14
+ "bs": 64,
15
+ "grad_accum": 4,
16
+ "lr": 0.0003,
17
+ "lora_lr": 0.0001,
18
+ "seed": 0,
19
+ "no_hybrid": false,
20
+ "no_cand_null": false,
21
+ "no_null": false,
22
+ "null": "factored",
23
+ "hard_only": false,
24
+ "mix": "full",
25
+ "eval_every": 4000,
26
+ "val_every": 1000,
27
+ "ckpt_every": 2000,
28
+ "eval_bs": 16,
29
+ "device": "auto",
30
+ "wandb": true,
31
+ "hf_repo": "guychuk/pcdm-runs",
32
+ "ckpt_upload": true,
33
+ "data": "data_v5",
34
+ "extra_data": "data_kb,data_wf,data_wf_hf,data_wf_long,data_wh,data_u",
35
+ "eval_cap": 1500,
36
+ "calib_sets": null,
37
+ "family_weights": "E:0.40,K:0.15,W:0.35,U:0.10",
38
+ "bucket_map": "data_wf_long=W,data_wh=W,data_u=U",
39
+ "null_aug": "W:0.20",
40
+ "smoke": false,
41
+ "eval_limit": 0,
42
+ "tap_layer": 26,
43
+ "zscore": true,
44
+ "eval_only": false,
45
+ "dump_logits": null,
46
+ "tower_d": 512,
47
+ "tower_layers": 2,
48
+ "tower_heads": 8,
49
+ "joint": false,
50
+ "listwise": false,
51
+ "extra_tap": 0,
52
+ "init_from": null,
53
+ "readout": "native",
54
+ "nc_head": "n3",
55
+ "nc_render": "letters_nonull",
56
+ "score_head": "choice",
57
+ "noul_head": "bern",
58
+ "qtype_filter": null,
59
+ "ordinal_smooth": 0.7,
60
+ "max_state": 1024,
61
+ "perm_lambda": 0.0,
62
+ "no_shuffle": false,
63
+ "zero_shot": false,
64
+ "shots": 0,
65
+ "cand_encoder": "backbone",
66
+ "head": "mlp",
67
+ "z_dim": 128,
68
+ "z_probes": 8,
69
+ "tiny_layers": 2,
70
+ "distill_alpha": 1.0,
71
+ "distill_beta": 0.0,
72
+ "distill_T": 2.0,
73
+ "delta_gamma": 0.0
74
+ },
75
+ "eval": {
76
+ "anli_test": {
77
+ "raw": {
78
+ "nll": 1.0757932427921113,
79
+ "brier": 0.6307436394770407,
80
+ "jsd": 0.2572920094851673,
81
+ "acc": 0.54375,
82
+ "acc_k": 0.54375,
83
+ "ece": 0.18585077961906793,
84
+ "sel_acc80": 0.572265625,
85
+ "coverage": 1.0,
86
+ "false_abstain": 0.0,
87
+ "null_recall": "n/a"
88
+ },
89
+ "scaled": {
90
+ "nll": 1.029092141141034,
91
+ "brier": 0.6156411190974276,
92
+ "jsd": 0.2564680245018555,
93
+ "acc": 0.54375,
94
+ "acc_k": 0.54375,
95
+ "ece": 0.1616302808281034,
96
+ "sel_acc80": 0.572265625,
97
+ "coverage": 1.0,
98
+ "false_abstain": 0.0,
99
+ "null_recall": "n/a"
100
+ }
101
+ },
102
+ "banking77_k": {
103
+ "raw": {
104
+ "nll": 0.8387672812179306,
105
+ "brier": 0.40366817680288347,
106
+ "jsd": 0.16477596631036817,
107
+ "acc": 0.7278933680104032,
108
+ "acc_k": 0.8718626155878467,
109
+ "ece": 0.1269044502406964,
110
+ "sel_acc80": 0.7996749288906948,
111
+ "conf_wrong": 0.11203585147247119,
112
+ "coverage": 0.6089076723016905,
113
+ "false_abstain": 0.11690885072655217,
114
+ "null_recall": 0.6568501920614597,
115
+ "auroc_null": 0.8789133262406189,
116
+ "auroc_conf": 0.8724135977145449
117
+ },
118
+ "scaled": {
119
+ "nll": 0.8239931862146926,
120
+ "brier": 0.4001605248196045,
121
+ "jsd": 0.16518185949943257,
122
+ "acc": 0.7298439531859557,
123
+ "acc_k": 0.8718626155878467,
124
+ "ece": 0.12038085420261124,
125
+ "sel_acc80": 0.8017066233238521,
126
+ "conf_wrong": 0.10627400768245839,
127
+ "coverage": 0.6059817945383614,
128
+ "false_abstain": 0.11889035667107001,
129
+ "null_recall": 0.6606914212548015,
130
+ "auroc_null": 0.878913326240619,
131
+ "auroc_conf": 0.8719218155093647
132
+ }
133
+ },
134
+ "banking77_test": {
135
+ "raw": {
136
+ "nll": 2.052295154509988,
137
+ "brier": 0.6301511272273523,
138
+ "jsd": 0.2807325439322352,
139
+ "acc": 0.5689206762028609,
140
+ "acc_k": 0.6404421326397919,
141
+ "ece": 0.1865983266862998,
142
+ "sel_acc80": 0.6468915075172694,
143
+ "coverage": 0.8299739921976593,
144
+ "false_abstain": 0.1700260078023407,
145
+ "null_recall": "n/a"
146
+ },
147
+ "scaled": {
148
+ "nll": 1.9304311410624564,
149
+ "brier": 0.6258988479081237,
150
+ "jsd": 0.2832575644475134,
151
+ "acc": 0.5679453836150845,
152
+ "acc_k": 0.6404421326397919,
153
+ "ece": 0.17139127035231674,
154
+ "sel_acc80": 0.6452661519707436,
155
+ "coverage": 0.8267230169050716,
156
+ "false_abstain": 0.17327698309492848,
157
+ "null_recall": "n/a"
158
+ }
159
+ },
160
+ "boolq_val": {
161
+ "raw": {
162
+ "nll": 0.39950305672739084,
163
+ "brier": 0.23047799103736755,
164
+ "jsd": 0.09612269180931764,
165
+ "acc": 0.8425076452599388,
166
+ "acc_k": 0.8440366972477065,
167
+ "ece": 0.049182466018819634,
168
+ "sel_acc80": 0.9063455657492355,
169
+ "coverage": 0.9969418960244648,
170
+ "false_abstain": 0.0030581039755351682,
171
+ "null_recall": "n/a"
172
+ },
173
+ "scaled": {
174
+ "nll": 0.3860076226788216,
175
+ "brier": 0.22776378177404963,
176
+ "jsd": 0.09772515374173538,
177
+ "acc": 0.8425076452599388,
178
+ "acc_k": 0.8440366972477065,
179
+ "ece": 0.036158523428330716,
180
+ "sel_acc80": 0.9063455657492355,
181
+ "coverage": 0.9969418960244648,
182
+ "false_abstain": 0.0030581039755351682,
183
+ "null_recall": "n/a"
184
+ }
185
+ },
186
+ "chaos_mnli": {
187
+ "raw": {
188
+ "nll": 1.2692985455080275,
189
+ "brier": 0.23347494851122355,
190
+ "jsd": 0.09021653349830507
191
+ },
192
+ "scaled": {
193
+ "nll": 1.188978150676804,
194
+ "brier": 0.21544714736486784,
195
+ "jsd": 0.0820067606796143
196
+ }
197
+ },
198
+ "clinc_heldout": {
199
+ "raw": {
200
+ "nll": 0.20331486754583272,
201
+ "brier": 0.09009287230064558,
202
+ "jsd": 0.04199571607272658,
203
+ "acc": 0.9365256124721604,
204
+ "acc_k": 0.9810690423162584,
205
+ "ece": 0.016201265172862793,
206
+ "sel_acc80": 0.9916550764951322,
207
+ "coverage": 0.9454342984409799,
208
+ "false_abstain": 0.05456570155902005,
209
+ "null_recall": "n/a"
210
+ },
211
+ "scaled": {
212
+ "nll": 0.19976629489598294,
213
+ "brier": 0.09080975891959425,
214
+ "jsd": 0.04278228699236031,
215
+ "acc": 0.9365256124721604,
216
+ "acc_k": 0.9810690423162584,
217
+ "ece": 0.017676310344103448,
218
+ "sel_acc80": 0.9902642559109874,
219
+ "coverage": 0.9443207126948775,
220
+ "false_abstain": 0.0556792873051225,
221
+ "null_recall": "n/a"
222
+ }
223
+ },
224
+ "clinc_k": {
225
+ "raw": {
226
+ "nll": 0.262570050582597,
227
+ "brier": 0.13814550219910562,
228
+ "jsd": 0.06344528720435406,
229
+ "acc": 0.9089478417266187,
230
+ "acc_k": 0.9789897183728208,
231
+ "ece": 0.011057701008729387,
232
+ "sel_acc80": 0.9674065748805845,
233
+ "conf_wrong": 0.03211216644052465,
234
+ "coverage": 0.5454136690647482,
235
+ "false_abstain": 0.04202056325435852,
236
+ "null_recall": 0.872003618272275,
237
+ "auroc_null": 0.9765736279790952,
238
+ "auroc_conf": 0.9762074740290501
239
+ },
240
+ "scaled": {
241
+ "nll": 0.2619289437378122,
242
+ "brier": 0.13735334589241208,
243
+ "jsd": 0.06378631528836894,
244
+ "acc": 0.9091726618705036,
245
+ "acc_k": 0.9789897183728208,
246
+ "ece": 0.00921771111864619,
247
+ "sel_acc80": 0.9676875526833381,
248
+ "conf_wrong": 0.03075531433740389,
249
+ "coverage": 0.5447392086330936,
250
+ "false_abstain": 0.0424675905230219,
251
+ "null_recall": 0.8729081863410222,
252
+ "auroc_null": 0.9765736279790951,
253
+ "auroc_conf": 0.9761183111952734
254
+ }
255
+ },
256
+ "clinc_oos": {
257
+ "raw": {
258
+ "nll": 0.8829221451581013,
259
+ "brier": 0.4253317641029075,
260
+ "jsd": 0.22263034211911978,
261
+ "acc": 0.7427427427427428,
262
+ "ece": 0.17314278620022072,
263
+ "sel_acc80": 0.775,
264
+ "conf_wrong": 0.044044044044044044,
265
+ "coverage": 0.25725725725725723,
266
+ "false_abstain": "n/a",
267
+ "null_recall": 0.7427427427427428
268
+ },
269
+ "scaled": {
270
+ "nll": 0.8829221442206856,
271
+ "brier": 0.4089999529102035,
272
+ "jsd": 0.2226303399234561,
273
+ "acc": 0.7607607607607607,
274
+ "ece": 0.185149793420826,
275
+ "sel_acc80": 0.8,
276
+ "conf_wrong": 0.042042042042042045,
277
+ "coverage": 0.2392392392392393,
278
+ "false_abstain": "n/a",
279
+ "null_recall": 0.7607607607607607
280
+ }
281
+ },
282
+ "clinc_test": {
283
+ "raw": {
284
+ "nll": 0.5477434777450255,
285
+ "brier": 0.21618169032775858,
286
+ "jsd": 0.10833343624727476,
287
+ "acc": 0.8502697841726619,
288
+ "acc_k": 0.8711780575539568,
289
+ "ece": 0.014199602169741796,
290
+ "sel_acc80": 0.9364990165776904,
291
+ "coverage": 0.9651528776978417,
292
+ "false_abstain": 0.03484712230215827,
293
+ "null_recall": "n/a"
294
+ },
295
+ "scaled": {
296
+ "nll": 0.5432309676259418,
297
+ "brier": 0.21901325867275181,
298
+ "jsd": 0.11382309839047851,
299
+ "acc": 0.8471223021582733,
300
+ "acc_k": 0.8711780575539568,
301
+ "ece": 0.01667415301764399,
302
+ "sel_acc80": 0.9342511941556617,
303
+ "coverage": 0.960431654676259,
304
+ "false_abstain": 0.039568345323741004,
305
+ "null_recall": "n/a"
306
+ }
307
+ },
308
+ "cse_banking77": {
309
+ "raw": {
310
+ "nll": 1.013373283262765,
311
+ "brier": 0.3896295995875706,
312
+ "jsd": 0.1607590626376104,
313
+ "acc": 0.7016666666666667,
314
+ "acc_k": 0.7608,
315
+ "ece": 0.14997901448110743,
316
+ "sel_acc80": 0.7683333333333333,
317
+ "conf_wrong": 0.118,
318
+ "coverage": 0.7903333333333333,
319
+ "false_abstain": 0.1108,
320
+ "null_recall": 0.704,
321
+ "auroc_null": 0.89648,
322
+ "auroc_conf": 0.8598568
323
+ },
324
+ "scaled": {
325
+ "nll": 0.9641588631983036,
326
+ "brier": 0.3855041258460942,
327
+ "jsd": 0.16056262866375498,
328
+ "acc": 0.702,
329
+ "acc_k": 0.7608,
330
+ "ece": 0.14450320369005204,
331
+ "sel_acc80": 0.7679166666666667,
332
+ "conf_wrong": 0.112,
333
+ "coverage": 0.789,
334
+ "false_abstain": 0.1116,
335
+ "null_recall": 0.708,
336
+ "auroc_null": 0.89648,
337
+ "auroc_conf": 0.8583504
338
+ },
339
+ "cse": {
340
+ "add_irr/p_irr": 0.000970285543377067,
341
+ "add_irr/dlo_top2": 0.22229034927480423,
342
+ "remove_gold/dp_null": 0.5173476701720647,
343
+ "dup/mass_err": 0.04696189815764079,
344
+ "dup/slot_gap": 0.20254385519145984,
345
+ "dup/acc_dedup": 0.852,
346
+ "reorder/max_dp": 0.09346072051447049,
347
+ "near_dup/cluster_mass_err": 0.0383210542500016
348
+ }
349
+ },
350
+ "cse_clinc": {
351
+ "raw": {
352
+ "nll": 0.27065101309237005,
353
+ "brier": 0.11199271731259507,
354
+ "jsd": 0.05256673519412974,
355
+ "acc": 0.8859800206682742,
356
+ "acc_k": 0.9152471956792688,
357
+ "ece": 0.013434116671178152,
358
+ "sel_acc80": 0.965992251399053,
359
+ "conf_wrong": 0.020161290322580645,
360
+ "coverage": 0.8157078883913194,
361
+ "false_abstain": 0.03946821769837973,
362
+ "null_recall": 0.8870967741935484,
363
+ "auroc_null": 0.9826153892544593,
364
+ "auroc_conf": 0.9730247463714703
365
+ },
366
+ "scaled": {
367
+ "nll": 0.2682153721308544,
368
+ "brier": 0.11078170689271184,
369
+ "jsd": 0.052865233459530446,
370
+ "acc": 0.8859800206682742,
371
+ "acc_k": 0.9152471956792688,
372
+ "ece": 0.01334414801801273,
373
+ "sel_acc80": 0.9655617735686612,
374
+ "conf_wrong": 0.018145161290322582,
375
+ "coverage": 0.8153634171546675,
376
+ "false_abstain": 0.039883672621520566,
377
+ "null_recall": 0.8870967741935484,
378
+ "auroc_null": 0.9826153892544593,
379
+ "auroc_conf": 0.9724878378921694
380
+ },
381
+ "cse": {
382
+ "add_irr/p_irr": 0.0004140982716602347,
383
+ "add_irr/dlo_top2": 0.21209305091759995,
384
+ "remove_gold/dp_null": 0.7651275384538431,
385
+ "dup/mass_err": 0.01625872549363192,
386
+ "dup/slot_gap": 0.16577797743548894,
387
+ "dup/acc_dedup": 0.9879032258064516,
388
+ "reorder/max_dp": 0.03674042043095158,
389
+ "near_dup/cluster_mass_err": 0.021475064313451826
390
+ }
391
+ },
392
+ "cse_hwu64": {
393
+ "raw": {
394
+ "nll": 0.47411119093999,
395
+ "brier": 0.21987119444109693,
396
+ "jsd": 0.10141276861597559,
397
+ "acc": 0.8153333333333334,
398
+ "acc_k": 0.8708,
399
+ "ece": 0.03724427296717967,
400
+ "sel_acc80": 0.8858333333333334,
401
+ "conf_wrong": 0.05,
402
+ "coverage": 0.8326666666666667,
403
+ "false_abstain": 0.0524,
404
+ "null_recall": 0.742,
405
+ "auroc_null": 0.9480720000000001,
406
+ "auroc_conf": 0.9206480000000001
407
+ },
408
+ "scaled": {
409
+ "nll": 0.4676719236027076,
410
+ "brier": 0.2173795947712966,
411
+ "jsd": 0.10222786247607862,
412
+ "acc": 0.814,
413
+ "acc_k": 0.8708,
414
+ "ece": 0.031223723649978607,
415
+ "sel_acc80": 0.8866666666666667,
416
+ "conf_wrong": 0.044,
417
+ "coverage": 0.8303333333333334,
418
+ "false_abstain": 0.0548,
419
+ "null_recall": 0.744,
420
+ "auroc_null": 0.9480720000000001,
421
+ "auroc_conf": 0.9197439999999999
422
+ },
423
+ "cse": {
424
+ "add_irr/p_irr": 0.006002017148867547,
425
+ "add_irr/dlo_top2": 0.1683097961744787,
426
+ "remove_gold/dp_null": 0.5516747428024682,
427
+ "dup/mass_err": 0.025967722096361542,
428
+ "dup/slot_gap": 0.17362533946861175,
429
+ "dup/acc_dedup": 0.96,
430
+ "reorder/max_dp": 0.05326351690868955,
431
+ "near_dup/cluster_mass_err": 0.03816221129417807
432
+ }
433
+ },
434
+ "cse_snli": {
435
+ "raw": {
436
+ "nll": 0.3785246814347904,
437
+ "brier": 0.15583854229184074,
438
+ "jsd": 0.07841771629000768,
439
+ "acc": 0.8942451420029895,
440
+ "acc_k": 0.9076286764705882,
441
+ "ece": 0.060462515327873466,
442
+ "sel_acc80": 0.932274638019617,
443
+ "conf_wrong": 0.032,
444
+ "coverage": 0.843796711509716,
445
+ "false_abstain": 0.0,
446
+ "null_recall": 0.836,
447
+ "auroc_null": 0.9999972426470588,
448
+ "auroc_conf": 0.9509972426470589
449
+ },
450
+ "scaled": {
451
+ "nll": 0.3844776583138124,
452
+ "brier": 0.1567990789175919,
453
+ "jsd": 0.08211913713577323,
454
+ "acc": 0.8942451420029895,
455
+ "acc_k": 0.9076286764705882,
456
+ "ece": 0.07534116962075416,
457
+ "sel_acc80": 0.932274638019617,
458
+ "conf_wrong": 0.032,
459
+ "coverage": 0.843796711509716,
460
+ "false_abstain": 0.0,
461
+ "null_recall": 0.836,
462
+ "auroc_null": 0.9999972426470588,
463
+ "auroc_conf": 0.9473409926470588
464
+ },
465
+ "cse": {
466
+ "add_irr/p_irr": 0.012405680973958624,
467
+ "add_irr/dlo_top2": 0.08886174408259696,
468
+ "remove_gold/dp_null": 0.7629259843477127,
469
+ "dup/mass_err": 0.04644630262634746,
470
+ "dup/slot_gap": 0.14531334546010474,
471
+ "dup/acc_dedup": 0.916,
472
+ "reorder/max_dp": 0.013759512538003036,
473
+ "near_dup/cluster_mass_err": 0.0401906658817924
474
+ }
475
+ },
476
+ "hwu64_test": {
477
+ "raw": {
478
+ "nll": 0.8544879862574016,
479
+ "brier": 0.3244043618126192,
480
+ "jsd": 0.16828853097185387,
481
+ "acc": 0.7732342007434945,
482
+ "acc_k": 0.8113382899628253,
483
+ "ece": 0.03379586631203675,
484
+ "sel_acc80": 0.8745644599303136,
485
+ "coverage": 0.9107806691449815,
486
+ "false_abstain": 0.08921933085501858,
487
+ "null_recall": "n/a"
488
+ },
489
+ "scaled": {
490
+ "nll": 0.8432026245083506,
491
+ "brier": 0.3286730332135295,
492
+ "jsd": 0.17449841753703743,
493
+ "acc": 0.7685873605947955,
494
+ "acc_k": 0.8113382899628253,
495
+ "ece": 0.03884527408500588,
496
+ "sel_acc80": 0.8722415795586528,
497
+ "coverage": 0.9014869888475836,
498
+ "false_abstain": 0.09851301115241635,
499
+ "null_recall": "n/a"
500
+ }
501
+ },
502
+ "ksweep_banking77": {
503
+ "raw": {
504
+ "nll": 0.9123393477006552,
505
+ "brier": 0.4099352651060742,
506
+ "jsd": 0.17143707584095402,
507
+ "acc": 0.7276785714285714,
508
+ "acc_k": 0.8607142857142858,
509
+ "ece": 0.12577233606151167,
510
+ "sel_acc80": 0.8020089285714286,
511
+ "conf_wrong": 0.09857142857142857,
512
+ "coverage": 0.6269642857142856,
513
+ "false_abstain": 0.095,
514
+ "null_recall": 0.6510714285714285,
515
+ "auroc_null": 0.8887267857142856,
516
+ "auroc_conf": 0.8845818877551019
517
+ },
518
+ "scaled": {
519
+ "nll": 0.8946802412802192,
520
+ "brier": 0.4049772048479024,
521
+ "jsd": 0.17224385846685789,
522
+ "acc": 0.7283928571428572,
523
+ "acc_k": 0.8607142857142858,
524
+ "ece": 0.11672890212652934,
525
+ "sel_acc80": 0.8040178571428571,
526
+ "conf_wrong": 0.09071428571428572,
527
+ "coverage": 0.6242857142857143,
528
+ "false_abstain": 0.09714285714285714,
529
+ "null_recall": 0.6542857142857142,
530
+ "auroc_null": 0.8887267857142856,
531
+ "auroc_conf": 0.8840718112244897
532
+ },
533
+ "ksweep": {
534
+ "K2/p_null_absent": 0.8811372525494516,
535
+ "K2/p_null_present": 0.15638932955197787,
536
+ "K2/acc": 0.975,
537
+ "K2/auroc": 0.9538375,
538
+ "K3/p_null_absent": 0.8291295790093136,
539
+ "K3/p_null_present": 0.11529847987123049,
540
+ "K3/acc": 0.9625,
541
+ "K3/auroc": 0.95703125,
542
+ "K5/p_null_absent": 0.7253072958214034,
543
+ "K5/p_null_present": 0.0856258794779842,
544
+ "K5/acc": 0.93,
545
+ "K5/auroc": 0.94483125,
546
+ "K10/p_null_absent": 0.6407146220555296,
547
+ "K10/p_null_present": 0.09918227566511632,
548
+ "K10/acc": 0.8925,
549
+ "K10/auroc": 0.9150375000000001,
550
+ "K20/p_null_absent": 0.4894634351064451,
551
+ "K20/p_null_present": 0.0933621416045935,
552
+ "K20/acc": 0.84,
553
+ "K20/auroc": 0.8745937500000001,
554
+ "K50/p_null_absent": 0.27896184352750425,
555
+ "K50/p_null_present": 0.05611725780938286,
556
+ "K50/acc": 0.735,
557
+ "K50/auroc": 0.809825,
558
+ "K77/p_null_absent": 0.39302291929721833,
559
+ "K77/p_null_present": 0.14252538830245612,
560
+ "K77/acc": 0.69,
561
+ "K77/auroc": 0.7562,
562
+ "p_null_absent_range": 0.6021754090219473
563
+ }
564
+ },
565
+ "ksweep_clinc": {
566
+ "raw": {
567
+ "nll": 0.33326689492364,
568
+ "brier": 0.1664739459458728,
569
+ "jsd": 0.0781240443304308,
570
+ "acc": 0.8875313283208021,
571
+ "acc_k": 0.956140350877193,
572
+ "ece": 0.014531420793543075,
573
+ "sel_acc80": 0.9541895066562255,
574
+ "conf_wrong": 0.03853383458646616,
575
+ "coverage": 0.5571741854636592,
576
+ "false_abstain": 0.039160401002506263,
577
+ "null_recall": 0.8464912280701754,
578
+ "auroc_null": 0.9752394378804153,
579
+ "auroc_conf": 0.9706104625599087
580
+ },
581
+ "scaled": {
582
+ "nll": 0.3334823865761253,
583
+ "brier": 0.16524676939838495,
584
+ "jsd": 0.07912781817534134,
585
+ "acc": 0.8894110275689223,
586
+ "acc_k": 0.956140350877193,
587
+ "ece": 0.014194780991022886,
588
+ "sel_acc80": 0.9549725920125294,
589
+ "conf_wrong": 0.03696741854636591,
590
+ "coverage": 0.5552944862155389,
591
+ "false_abstain": 0.039160401002506263,
592
+ "null_recall": 0.850250626566416,
593
+ "auroc_null": 0.9752394378804151,
594
+ "auroc_conf": 0.9697932958963825
595
+ },
596
+ "ksweep": {
597
+ "K2/p_null_absent": 0.9124401700321146,
598
+ "K2/p_null_present": 0.03731175375713476,
599
+ "K2/acc": 1.0,
600
+ "K2/auroc": 0.9960301756898512,
601
+ "K3/p_null_absent": 0.8935738604922515,
602
+ "K3/p_null_present": 0.03285068903447748,
603
+ "K3/acc": 0.9974937343358395,
604
+ "K3/auroc": 0.9916143742815686,
605
+ "K5/p_null_absent": 0.8634682853127313,
606
+ "K5/p_null_present": 0.038481267121283975,
607
+ "K5/acc": 0.9949874686716792,
608
+ "K5/auroc": 0.989679713067129,
609
+ "K10/p_null_absent": 0.8389577590959627,
610
+ "K10/p_null_present": 0.0566416007986982,
611
+ "K10/acc": 0.9874686716791979,
612
+ "K10/auroc": 0.9814197147002844,
613
+ "K20/p_null_absent": 0.7870685033885911,
614
+ "K20/p_null_present": 0.0636121057681325,
615
+ "K20/acc": 0.9573934837092731,
616
+ "K20/auroc": 0.971645906746817,
617
+ "K50/p_null_absent": 0.7133180369491127,
618
+ "K50/p_null_present": 0.08933361823714915,
619
+ "K50/acc": 0.9448621553884712,
620
+ "K50/auroc": 0.9461686798449758,
621
+ "K100/p_null_absent": 0.4153825638495983,
622
+ "K100/p_null_present": 0.05320865131507892,
623
+ "K100/acc": 0.9147869674185464,
624
+ "K100/auroc": 0.9146739028021182,
625
+ "K150/p_null_absent": 0.8475411443417906,
626
+ "K150/p_null_present": 0.054351499962986896,
627
+ "K150/acc": 0.8521303258145363,
628
+ "K150/auroc": 0.9914259332541882,
629
+ "p_null_absent_range": 0.4970576061825163
630
+ }
631
+ },
632
+ "mmlu_cf": {
633
+ "raw": {
634
+ "nll": 1.7757699338936928,
635
+ "brier": 0.6361527077025765,
636
+ "jsd": 0.3573300858114095,
637
+ "acc": 0.43933333333333335,
638
+ "acc_k": 0.4513333333333333,
639
+ "ece": 0.059149413963158926,
640
+ "sel_acc80": 0.495,
641
+ "coverage": 0.9806666666666667,
642
+ "false_abstain": 0.019333333333333334,
643
+ "null_recall": "n/a"
644
+ },
645
+ "scaled": {
646
+ "nll": 1.7661442234094802,
647
+ "brier": 0.6354676987043012,
648
+ "jsd": 0.363000626599059,
649
+ "acc": 0.43933333333333335,
650
+ "acc_k": 0.4513333333333333,
651
+ "ece": 0.06224065233767032,
652
+ "sel_acc80": 0.49416666666666664,
653
+ "coverage": 0.98,
654
+ "false_abstain": 0.02,
655
+ "null_recall": "n/a"
656
+ },
657
+ "cf": {
658
+ "orig/acc": 0.472,
659
+ "orig/among_k": 0.472,
660
+ "remove3/acc": 0.476,
661
+ "remove3/among_k": 0.536,
662
+ "add3_unrel/acc": 0.444,
663
+ "add3_unrel/among_k": 0.452,
664
+ "replace_hard/acc": 0.452,
665
+ "replace_hard/among_k": 0.456,
666
+ "add_neardup/acc": 0.292,
667
+ "add_neardup/among_k": 0.292,
668
+ "reorder/acc": 0.5,
669
+ "reorder/among_k": 0.5,
670
+ "add3_unrel/dlo_top2": 0.19286915780863595,
671
+ "reorder/max_dp": 0.1489373840354383,
672
+ "add_neardup/slot_gap": 0.13364317205362022,
673
+ "add_neardup/cluster_mass_err": 0.1036915674302727
674
+ }
675
+ },
676
+ "mmlu_cf_teacher": {
677
+ "raw": {
678
+ "nll": 1.7413541804238846,
679
+ "brier": 0.6266079920746009,
680
+ "jsd": 0.3541206344382742,
681
+ "acc": 0.4297222222222222,
682
+ "acc_k": 0.44305555555555554,
683
+ "ece": 0.03678819169600805,
684
+ "sel_acc80": 0.48420138888888886,
685
+ "coverage": 0.9794444444444445,
686
+ "false_abstain": 0.020555555555555556,
687
+ "null_recall": "n/a"
688
+ },
689
+ "scaled": {
690
+ "nll": 1.7379008911888407,
691
+ "brier": 0.6271273947648627,
692
+ "jsd": 0.36024415790391084,
693
+ "acc": 0.4297222222222222,
694
+ "acc_k": 0.44305555555555554,
695
+ "ece": 0.045424607180886804,
696
+ "sel_acc80": 0.4852430555555556,
697
+ "coverage": 0.9791666666666666,
698
+ "false_abstain": 0.020833333333333332,
699
+ "null_recall": "n/a"
700
+ },
701
+ "cf": {
702
+ "orig/acc": 0.4558333333333333,
703
+ "orig/among_k": 0.45916666666666667,
704
+ "orig/kl_teacher": 0.44995319883388635,
705
+ "remove3/acc": 0.4766666666666667,
706
+ "remove3/among_k": 0.5366666666666666,
707
+ "remove3/kl_teacher": 1.2494058071732164,
708
+ "add3_unrel/acc": 0.4575,
709
+ "add3_unrel/among_k": 0.4633333333333333,
710
+ "add3_unrel/kl_teacher": 0.5880642448495372,
711
+ "replace_hard/acc": 0.45166666666666666,
712
+ "replace_hard/among_k": 0.4583333333333333,
713
+ "replace_hard/kl_teacher": 0.48544187695503743,
714
+ "add_neardup/acc": 0.28,
715
+ "add_neardup/among_k": 0.28,
716
+ "add_neardup/kl_teacher": 0.42830956656334196,
717
+ "reorder/acc": 0.45666666666666667,
718
+ "reorder/among_k": 0.4608333333333333,
719
+ "reorder/kl_teacher": 0.44567269445063523,
720
+ "remove3/delta_mae": 1.2099147179202685,
721
+ "remove3/delta_teacher_abs": 1.135840148062828,
722
+ "add3_unrel/delta_mae": 0.7671681344547404,
723
+ "add3_unrel/delta_teacher_abs": 0.7390631803199154,
724
+ "replace_hard/delta_mae": 0.7565496808691786,
725
+ "replace_hard/delta_teacher_abs": 0.6681882120617108,
726
+ "add_neardup/delta_mae": 0.7833307280802294,
727
+ "add_neardup/delta_teacher_abs": 0.7305883395327554,
728
+ "reorder/delta_mae": 1.059772158733795,
729
+ "reorder/delta_teacher_abs": 0.7763704285195197,
730
+ "add3_unrel/dlo_top2": 0.17924617675632248,
731
+ "reorder/max_dp": 0.1370161229314774,
732
+ "add_neardup/slot_gap": 0.13539707640166548,
733
+ "add_neardup/cluster_mass_err": 0.10971532651535627
734
+ }
735
+ },
736
+ "mmlu_choicesonly": {
737
+ "raw": {
738
+ "nll": 2.039828592927867,
739
+ "brier": 0.8168081995898367,
740
+ "jsd": 0.4534946954898565,
741
+ "acc": 0.29,
742
+ "acc_k": 0.29083333333333333,
743
+ "ece": 0.04304502696419755,
744
+ "sel_acc80": 0.3229166666666667,
745
+ "coverage": 0.9966666666666667,
746
+ "false_abstain": 0.0033333333333333335,
747
+ "null_recall": "n/a"
748
+ },
749
+ "scaled": {
750
+ "nll": 2.0257116311704886,
751
+ "brier": 0.8150993429201715,
752
+ "jsd": 0.4567105766310495,
753
+ "acc": 0.29,
754
+ "acc_k": 0.29083333333333333,
755
+ "ece": 0.03703421314557393,
756
+ "sel_acc80": 0.32395833333333335,
757
+ "coverage": 0.9966666666666667,
758
+ "false_abstain": 0.0033333333333333335,
759
+ "null_recall": "n/a"
760
+ }
761
+ },
762
+ "mmlu_pro": {
763
+ "raw": {
764
+ "nll": 1.6070274236514384,
765
+ "brier": 0.6767720553604334,
766
+ "jsd": 0.3678978835677458,
767
+ "acc": 0.455,
768
+ "acc_k": 0.4583333333333333,
769
+ "ece": 0.04268285360187291,
770
+ "sel_acc80": 0.5166666666666667,
771
+ "coverage": 0.9958333333333333,
772
+ "false_abstain": 0.004166666666666667,
773
+ "null_recall": "n/a"
774
+ },
775
+ "scaled": {
776
+ "nll": 1.6083517592633751,
777
+ "brier": 0.6793089765777306,
778
+ "jsd": 0.3746115181488698,
779
+ "acc": 0.455,
780
+ "acc_k": 0.4583333333333333,
781
+ "ece": 0.05830767397458354,
782
+ "sel_acc80": 0.5166666666666667,
783
+ "coverage": 0.9958333333333333,
784
+ "false_abstain": 0.004166666666666667,
785
+ "null_recall": "n/a"
786
+ }
787
+ },
788
+ "mmlu_shuffledq": {
789
+ "raw": {
790
+ "nll": 2.0994824466839295,
791
+ "brier": 0.839300334664014,
792
+ "jsd": 0.4620919566125671,
793
+ "acc": 0.2633333333333333,
794
+ "acc_k": 0.265,
795
+ "ece": 0.06165247778718671,
796
+ "sel_acc80": 0.29270833333333335,
797
+ "coverage": 0.9933333333333333,
798
+ "false_abstain": 0.006666666666666667,
799
+ "null_recall": "n/a"
800
+ },
801
+ "scaled": {
802
+ "nll": 2.0802900159544735,
803
+ "brier": 0.8357595643299349,
804
+ "jsd": 0.46472209676556786,
805
+ "acc": 0.2633333333333333,
806
+ "acc_k": 0.265,
807
+ "ece": 0.03918854349603254,
808
+ "sel_acc80": 0.29270833333333335,
809
+ "coverage": 0.9933333333333333,
810
+ "false_abstain": 0.006666666666666667,
811
+ "null_recall": "n/a"
812
+ }
813
+ },
814
+ "mnli_val": {
815
+ "raw": {
816
+ "nll": 0.3773646737665917,
817
+ "brier": 0.20883140960217866,
818
+ "jsd": 0.09443824642830204,
819
+ "acc": 0.8605442176870748,
820
+ "acc_k": 0.8605442176870748,
821
+ "ece": 0.020842784633036352,
822
+ "sel_acc80": 0.9200680272108843,
823
+ "coverage": 1.0,
824
+ "false_abstain": 0.0,
825
+ "null_recall": "n/a"
826
+ },
827
+ "scaled": {
828
+ "nll": 0.3750955821527883,
829
+ "brier": 0.20850910256182303,
830
+ "jsd": 0.09836140994889432,
831
+ "acc": 0.8605442176870748,
832
+ "acc_k": 0.8605442176870748,
833
+ "ece": 0.009489905398313684,
834
+ "sel_acc80": 0.9200680272108843,
835
+ "coverage": 1.0,
836
+ "false_abstain": 0.0,
837
+ "null_recall": "n/a"
838
+ }
839
+ },
840
+ "ng20_test": {
841
+ "raw": {
842
+ "nll": 1.487761037323019,
843
+ "brier": 0.5656505449790027,
844
+ "jsd": 0.26489029125897295,
845
+ "acc": 0.5913875598086125,
846
+ "acc_k": 0.6397812713602188,
847
+ "ece": 0.13154189228726346,
848
+ "sel_acc80": 0.6712235133287765,
849
+ "coverage": 0.8489405331510594,
850
+ "false_abstain": 0.15105946684894053,
851
+ "null_recall": "n/a"
852
+ },
853
+ "scaled": {
854
+ "nll": 1.4293305916943673,
855
+ "brier": 0.5619083493741223,
856
+ "jsd": 0.2683506922940156,
857
+ "acc": 0.5882433356117567,
858
+ "acc_k": 0.6397812713602188,
859
+ "ece": 0.11515617097841906,
860
+ "sel_acc80": 0.6669514695830485,
861
+ "coverage": 0.8414217361585783,
862
+ "false_abstain": 0.15857826384142173,
863
+ "null_recall": "n/a"
864
+ }
865
+ },
866
+ "null_irrq_intent": {
867
+ "raw": {
868
+ "nll": 0.2228418101184203,
869
+ "brier": 0.1143710591296055,
870
+ "jsd": 0.06114704381982695,
871
+ "acc": 0.9266936299292214,
872
+ "acc_k": 0.9816326530612245,
873
+ "ece": 0.031551026387933015,
874
+ "sel_acc80": 0.981680353758686,
875
+ "conf_wrong": 0.006012024048096192,
876
+ "coverage": 0.5197168857431749,
877
+ "false_abstain": 0.04387755102040816,
878
+ "null_recall": 0.908817635270541,
879
+ "auroc_null": 0.9819393889820457,
880
+ "auroc_conf": 0.9848083922947937
881
+ },
882
+ "scaled": {
883
+ "nll": 0.2227184423843362,
884
+ "brier": 0.11267167675609457,
885
+ "jsd": 0.06145260385603397,
886
+ "acc": 0.9277047522750252,
887
+ "acc_k": 0.9816326530612245,
888
+ "ece": 0.036607042256332135,
889
+ "sel_acc80": 0.9823120656980417,
890
+ "conf_wrong": 0.004008016032064128,
891
+ "coverage": 0.5182002022244692,
892
+ "false_abstain": 0.044897959183673466,
893
+ "null_recall": 0.9108216432865731,
894
+ "auroc_null": 0.9819393889820457,
895
+ "auroc_conf": 0.9853574495930636
896
+ }
897
+ },
898
+ "null_irrq_squad": {
899
+ "raw": {
900
+ "nll": 0.11958802708388418,
901
+ "brier": 0.06691821229442936,
902
+ "jsd": 0.031051279929708648,
903
+ "acc": 0.9555,
904
+ "acc_k": 0.995,
905
+ "ece": 0.006837947338819474,
906
+ "sel_acc80": 0.993125,
907
+ "conf_wrong": 0.015,
908
+ "coverage": 0.5355,
909
+ "false_abstain": 0.007,
910
+ "null_recall": 0.922,
911
+ "auroc_null": 0.9963829999999999,
912
+ "auroc_conf": 0.996046
913
+ },
914
+ "scaled": {
915
+ "nll": 0.11969051312469098,
916
+ "brier": 0.06648639497090965,
917
+ "jsd": 0.031141734225993013,
918
+ "acc": 0.9555,
919
+ "acc_k": 0.995,
920
+ "ece": 0.009631279647350293,
921
+ "sel_acc80": 0.994375,
922
+ "conf_wrong": 0.014,
923
+ "coverage": 0.5355,
924
+ "false_abstain": 0.007,
925
+ "null_recall": 0.922,
926
+ "auroc_null": 0.996383,
927
+ "auroc_conf": 0.996008
928
+ }
929
+ },
930
+ "null_nearmiss_banking77": {
931
+ "raw": {
932
+ "nll": 0.9175728007055943,
933
+ "brier": 0.4502252277332544,
934
+ "jsd": 0.1821161578591737,
935
+ "acc": 0.704,
936
+ "acc_k": 0.882,
937
+ "ece": 0.14530414193868638,
938
+ "sel_acc80": 0.75875,
939
+ "conf_wrong": 0.148,
940
+ "coverage": 0.6639999999999999,
941
+ "false_abstain": 0.092,
942
+ "null_recall": 0.58,
943
+ "auroc_null": 0.8760680000000001,
944
+ "auroc_conf": 0.8604099999999999
945
+ },
946
+ "scaled": {
947
+ "nll": 0.9055373980443426,
948
+ "brier": 0.44583399808413937,
949
+ "jsd": 0.18265736516534373,
950
+ "acc": 0.706,
951
+ "acc_k": 0.882,
952
+ "ece": 0.13480573588609693,
953
+ "sel_acc80": 0.7625,
954
+ "conf_wrong": 0.136,
955
+ "coverage": 0.6619999999999999,
956
+ "false_abstain": 0.092,
957
+ "null_recall": 0.584,
958
+ "auroc_null": 0.8760680000000001,
959
+ "auroc_conf": 0.859052
960
+ }
961
+ },
962
+ "null_nearmiss_clinc": {
963
+ "raw": {
964
+ "nll": 0.5506594233951506,
965
+ "brier": 0.2877703155362219,
966
+ "jsd": 0.11687679987648988,
967
+ "acc": 0.8135764944275583,
968
+ "acc_k": 0.9494949494949495,
969
+ "ece": 0.08668260612990229,
970
+ "sel_acc80": 0.8708860759493671,
971
+ "conf_wrong": 0.13414634146341464,
972
+ "coverage": 0.6322188449848024,
973
+ "false_abstain": 0.03636363636363636,
974
+ "null_recall": 0.7012195121951219,
975
+ "auroc_null": 0.9421162847992116,
976
+ "auroc_conf": 0.9255440584708877
977
+ },
978
+ "scaled": {
979
+ "nll": 0.5473392438642083,
980
+ "brier": 0.28632701776783964,
981
+ "jsd": 0.1173640732811478,
982
+ "acc": 0.8135764944275583,
983
+ "acc_k": 0.9494949494949495,
984
+ "ece": 0.08807264780201081,
985
+ "sel_acc80": 0.8734177215189873,
986
+ "conf_wrong": 0.13008130081300814,
987
+ "coverage": 0.6322188449848024,
988
+ "false_abstain": 0.03636363636363636,
989
+ "null_recall": 0.7012195121951219,
990
+ "auroc_null": 0.9421162847992116,
991
+ "auroc_conf": 0.9237004188223701
992
+ }
993
+ },
994
+ "null_nearmiss_hwu64": {
995
+ "raw": {
996
+ "nll": 0.6930345077628086,
997
+ "brier": 0.37550991179408877,
998
+ "jsd": 0.1627737776269167,
999
+ "acc": 0.734,
1000
+ "acc_k": 0.924,
1001
+ "ece": 0.08103191289305686,
1002
+ "sel_acc80": 0.80125,
1003
+ "conf_wrong": 0.092,
1004
+ "coverage": 0.6799999999999999,
1005
+ "false_abstain": 0.058,
1006
+ "null_recall": 0.582,
1007
+ "auroc_null": 0.914212,
1008
+ "auroc_conf": 0.8997919999999999
1009
+ },
1010
+ "scaled": {
1011
+ "nll": 0.691222061632767,
1012
+ "brier": 0.3714586770346436,
1013
+ "jsd": 0.16380177240954913,
1014
+ "acc": 0.736,
1015
+ "acc_k": 0.924,
1016
+ "ece": 0.07761898154020311,
1017
+ "sel_acc80": 0.80625,
1018
+ "conf_wrong": 0.092,
1019
+ "coverage": 0.6779999999999999,
1020
+ "false_abstain": 0.058,
1021
+ "null_recall": 0.586,
1022
+ "auroc_null": 0.914212,
1023
+ "auroc_conf": 0.898608
1024
+ }
1025
+ },
1026
+ "snli_null": {
1027
+ "raw": {
1028
+ "nll": 0.3419294368754166,
1029
+ "brier": 0.1853403799127337,
1030
+ "jsd": 0.08550839024504887,
1031
+ "acc": 0.8741230707556624,
1032
+ "acc_k": 0.9477732793522268,
1033
+ "ece": 0.019044966377746518,
1034
+ "sel_acc80": 0.9418837675350702,
1035
+ "conf_wrong": 0.037316395394998016,
1036
+ "coverage": 0.5742633794347565,
1037
+ "false_abstain": 0.023076923076923078,
1038
+ "null_recall": 0.8205637157602224,
1039
+ "auroc_null": 0.9805860882395012,
1040
+ "auroc_conf": 0.9665643940063613
1041
+ },
1042
+ "scaled": {
1043
+ "nll": 0.343350635842443,
1044
+ "brier": 0.18461307863262783,
1045
+ "jsd": 0.08667314826430747,
1046
+ "acc": 0.8751252756063339,
1047
+ "acc_k": 0.9477732793522268,
1048
+ "ece": 0.019226861922385895,
1049
+ "sel_acc80": 0.9418837675350702,
1050
+ "conf_wrong": 0.03652242953552997,
1051
+ "coverage": 0.573261174584085,
1052
+ "false_abstain": 0.023076923076923078,
1053
+ "null_recall": 0.8225486304088924,
1054
+ "auroc_null": 0.9805860882395012,
1055
+ "auroc_conf": 0.9652861732613514
1056
+ }
1057
+ },
1058
+ "snli_test": {
1059
+ "raw": {
1060
+ "nll": 0.25593494718671267,
1061
+ "brier": 0.14036088145874998,
1062
+ "jsd": 0.0665975840595549,
1063
+ "acc": 0.9089997995590299,
1064
+ "acc_k": 0.9089997995590299,
1065
+ "ece": 0.008833283340077942,
1066
+ "sel_acc80": 0.9611723446893787,
1067
+ "coverage": 1.0,
1068
+ "false_abstain": 0.0,
1069
+ "null_recall": "n/a"
1070
+ },
1071
+ "scaled": {
1072
+ "nll": 0.25976074002227,
1073
+ "brier": 0.14143341682032448,
1074
+ "jsd": 0.07073793640612969,
1075
+ "acc": 0.9089997995590299,
1076
+ "acc_k": 0.9089997995590299,
1077
+ "ece": 0.023527867845648823,
1078
+ "sel_acc80": 0.9614228456913828,
1079
+ "coverage": 1.0,
1080
+ "false_abstain": 0.0,
1081
+ "null_recall": "n/a"
1082
+ }
1083
+ },
1084
+ "snli_test_hyponly": {
1085
+ "raw": {
1086
+ "nll": 1.7751328278047886,
1087
+ "brier": 0.9288499904530326,
1088
+ "jsd": 0.350175634798463,
1089
+ "acc": 0.3742232912407296,
1090
+ "acc_k": 0.3742232912407296,
1091
+ "ece": 0.42920332531170124,
1092
+ "sel_acc80": 0.3905310621242485,
1093
+ "coverage": 1.0,
1094
+ "false_abstain": 0.0,
1095
+ "null_recall": "n/a"
1096
+ },
1097
+ "scaled": {
1098
+ "nll": 1.6366306739667331,
1099
+ "brier": 0.894582538842687,
1100
+ "jsd": 0.34275924283904896,
1101
+ "acc": 0.3742232912407296,
1102
+ "acc_k": 0.3742232912407296,
1103
+ "ece": 0.4047492533521964,
1104
+ "sel_acc80": 0.39203406813627256,
1105
+ "coverage": 1.0,
1106
+ "false_abstain": 0.0,
1107
+ "null_recall": "n/a"
1108
+ }
1109
+ },
1110
+ "snli_test_paraphrase": {
1111
+ "raw": {
1112
+ "nll": 0.32587851940146534,
1113
+ "brier": 0.17727392068849784,
1114
+ "jsd": 0.0913833239073259,
1115
+ "acc": 0.8877530567247945,
1116
+ "acc_k": 0.8877530567247945,
1117
+ "ece": 0.04598980332023037,
1118
+ "sel_acc80": 0.9438877755511023,
1119
+ "coverage": 1.0,
1120
+ "false_abstain": 0.0,
1121
+ "null_recall": "n/a"
1122
+ },
1123
+ "scaled": {
1124
+ "nll": 0.33854533963891326,
1125
+ "brier": 0.18208532264372873,
1126
+ "jsd": 0.09776032407399843,
1127
+ "acc": 0.8877530567247945,
1128
+ "acc_k": 0.8877530567247945,
1129
+ "ece": 0.06781542564016661,
1130
+ "sel_acc80": 0.9443887775551102,
1131
+ "coverage": 1.0,
1132
+ "false_abstain": 0.0,
1133
+ "null_recall": "n/a"
1134
+ }
1135
+ },
1136
+ "snli_test_qpara": {
1137
+ "raw": {
1138
+ "nll": 0.24387772870151897,
1139
+ "brier": 0.13415882157921932,
1140
+ "jsd": 0.061499692705830125,
1141
+ "acc": 0.9112337011033099,
1142
+ "acc_k": 0.9112337011033099,
1143
+ "ece": 0.01218783982479228,
1144
+ "sel_acc80": 0.9655388471177945,
1145
+ "coverage": 1.0,
1146
+ "false_abstain": 0.0,
1147
+ "null_recall": "n/a"
1148
+ },
1149
+ "scaled": {
1150
+ "nll": 0.24449059889889446,
1151
+ "brier": 0.13416361658852846,
1152
+ "jsd": 0.06481072262916734,
1153
+ "acc": 0.9112337011033099,
1154
+ "acc_k": 0.9112337011033099,
1155
+ "ece": 0.01282951259684777,
1156
+ "sel_acc80": 0.9655388471177945,
1157
+ "coverage": 1.0,
1158
+ "false_abstain": 0.0,
1159
+ "null_recall": "n/a"
1160
+ }
1161
+ },
1162
+ "snli_test_soft": {
1163
+ "raw": {
1164
+ "nll": 0.5597517215152261,
1165
+ "brier": 0.09879588537311783,
1166
+ "jsd": 0.05306827216224841,
1167
+ "acc": 0.9089997995590299,
1168
+ "acc_k": 0.9089997995590299,
1169
+ "ece": 0.008833283340077942,
1170
+ "sel_acc80": 0.9611723446893787,
1171
+ "coverage": 1.0,
1172
+ "false_abstain": 0.0,
1173
+ "null_recall": "n/a"
1174
+ },
1175
+ "scaled": {
1176
+ "nll": 0.5300098613125277,
1177
+ "brier": 0.09536717726609922,
1178
+ "jsd": 0.053384305599445174,
1179
+ "acc": 0.9089997995590299,
1180
+ "acc_k": 0.9089997995590299,
1181
+ "ece": 0.023527867845648823,
1182
+ "sel_acc80": 0.9614228456913828,
1183
+ "coverage": 1.0,
1184
+ "false_abstain": 0.0,
1185
+ "null_recall": "n/a"
1186
+ }
1187
+ },
1188
+ "squad_null": {
1189
+ "raw": {
1190
+ "nll": 0.0653928564491927,
1191
+ "brier": 0.03340959325302636,
1192
+ "jsd": 0.014394054113564767,
1193
+ "acc": 0.9775,
1194
+ "acc_k": 0.997,
1195
+ "ece": 0.007179727219045159,
1196
+ "sel_acc80": 0.9975,
1197
+ "conf_wrong": 0.0115,
1198
+ "coverage": 0.507,
1199
+ "false_abstain": 0.0145,
1200
+ "null_recall": 0.9715,
1201
+ "auroc_null": 0.997093,
1202
+ "auroc_conf": 0.9969047500000001
1203
+ },
1204
+ "scaled": {
1205
+ "nll": 0.06562872611683704,
1206
+ "brier": 0.03343094242134052,
1207
+ "jsd": 0.014530513671621156,
1208
+ "acc": 0.97775,
1209
+ "acc_k": 0.997,
1210
+ "ece": 0.007410270601511033,
1211
+ "sel_acc80": 0.9975,
1212
+ "conf_wrong": 0.0105,
1213
+ "coverage": 0.50675,
1214
+ "false_abstain": 0.0145,
1215
+ "null_recall": 0.972,
1216
+ "auroc_null": 0.997093,
1217
+ "auroc_conf": 0.99687025
1218
+ }
1219
+ },
1220
+ "trec_coarse": {
1221
+ "raw": {
1222
+ "nll": 1.001176772570978,
1223
+ "brier": 0.45645891235995145,
1224
+ "jsd": 0.23270860839413132,
1225
+ "acc": 0.708,
1226
+ "acc_k": 0.744,
1227
+ "ece": 0.07937431913614273,
1228
+ "sel_acc80": 0.7725,
1229
+ "coverage": 0.91,
1230
+ "false_abstain": 0.09,
1231
+ "null_recall": "n/a"
1232
+ },
1233
+ "scaled": {
1234
+ "nll": 0.9972639353463436,
1235
+ "brier": 0.46413852182874643,
1236
+ "jsd": 0.24018308427283502,
1237
+ "acc": 0.706,
1238
+ "acc_k": 0.744,
1239
+ "ece": 0.10051598793268204,
1240
+ "sel_acc80": 0.765,
1241
+ "coverage": 0.908,
1242
+ "false_abstain": 0.092,
1243
+ "null_recall": "n/a"
1244
+ }
1245
+ },
1246
+ "trec_fine": {
1247
+ "raw": {
1248
+ "nll": 1.9270376611515554,
1249
+ "brier": 0.7866914685130448,
1250
+ "jsd": 0.3710943789415943,
1251
+ "acc": 0.42,
1252
+ "acc_k": 0.51,
1253
+ "ece": 0.20852845552563673,
1254
+ "sel_acc80": 0.44,
1255
+ "coverage": 0.752,
1256
+ "false_abstain": 0.248,
1257
+ "null_recall": "n/a"
1258
+ },
1259
+ "scaled": {
1260
+ "nll": 1.890897607132656,
1261
+ "brier": 0.7828980197231503,
1262
+ "jsd": 0.37848747499812385,
1263
+ "acc": 0.414,
1264
+ "acc_k": 0.51,
1265
+ "ece": 0.18710149055719377,
1266
+ "sel_acc80": 0.4275,
1267
+ "coverage": 0.734,
1268
+ "false_abstain": 0.266,
1269
+ "null_recall": "n/a"
1270
+ }
1271
+ },
1272
+ "unli_test": {
1273
+ "raw": {
1274
+ "nll": 0.5144665216131059,
1275
+ "brier": 0.11676567490046662,
1276
+ "jsd": 0.04409186280928513,
1277
+ "auroc_null": 0.932519288829252,
1278
+ "auroc_conf": 0.932519288829252
1279
+ },
1280
+ "scaled": {
1281
+ "nll": 0.5144665214847596,
1282
+ "brier": 0.11676567514171127,
1283
+ "jsd": 0.044091862724640454,
1284
+ "auroc_null": 0.932519288829252,
1285
+ "auroc_conf": 0.932519288829252
1286
+ }
1287
+ },
1288
+ "data_kb_val": {
1289
+ "raw": {
1290
+ "nll": 0.7956040995278478,
1291
+ "brier": 0.4090308878761947,
1292
+ "jsd": 0.2035840258631193,
1293
+ "acc": 0.6826666666666666,
1294
+ "acc_k": 0.6659634317862166,
1295
+ "ece": 0.021778101732333507,
1296
+ "sel_acc80": 0.7691666666666667,
1297
+ "conf_wrong": 0.0,
1298
+ "coverage": 0.948,
1299
+ "false_abstain": 0.0007032348804500703,
1300
+ "null_recall": 0.9871794871794872,
1301
+ "auroc_null": 0.9999819683363987,
1302
+ "auroc_conf": 0.9962313823073317
1303
+ },
1304
+ "scaled": {
1305
+ "nll": 0.7956797234491352,
1306
+ "brier": 0.409640835609177,
1307
+ "jsd": 0.2081642635211449,
1308
+ "acc": 0.6826666666666666,
1309
+ "acc_k": 0.6659634317862166,
1310
+ "ece": 0.023031593660513546,
1311
+ "sel_acc80": 0.7683333333333333,
1312
+ "conf_wrong": 0.0,
1313
+ "coverage": 0.948,
1314
+ "false_abstain": 0.0007032348804500703,
1315
+ "null_recall": 0.9871794871794872,
1316
+ "auroc_null": 0.9999819683363987,
1317
+ "auroc_conf": 0.9961502398211259
1318
+ }
1319
+ },
1320
+ "truthfulqa_mc1": {
1321
+ "raw": {
1322
+ "nll": 2.1805148426121623,
1323
+ "brier": 0.8215843259734116,
1324
+ "jsd": 0.380775613960407,
1325
+ "acc": 0.408812729498164,
1326
+ "acc_k": 0.46511627906976744,
1327
+ "ece": 0.1856523987676406,
1328
+ "sel_acc80": 0.44954128440366975,
1329
+ "coverage": 0.8629130966952264,
1330
+ "false_abstain": 0.13708690330477355,
1331
+ "null_recall": "n/a"
1332
+ },
1333
+ "scaled": {
1334
+ "nll": 2.1274534999375962,
1335
+ "brier": 0.8176668121988437,
1336
+ "jsd": 0.383300010835186,
1337
+ "acc": 0.40758873929008566,
1338
+ "acc_k": 0.46511627906976744,
1339
+ "ece": 0.1614636940245529,
1340
+ "sel_acc80": 0.44801223241590216,
1341
+ "coverage": 0.8604651162790697,
1342
+ "false_abstain": 0.13953488372093023,
1343
+ "null_recall": "n/a"
1344
+ }
1345
+ },
1346
+ "data_wf_val": {
1347
+ "raw": {
1348
+ "nll": 0.338919784759892,
1349
+ "brier": 0.18760861684389032,
1350
+ "jsd": 0.09645376573672314,
1351
+ "acc": 0.878,
1352
+ "acc_k": 0.8759372869802318,
1353
+ "ece": 0.05888278289635974,
1354
+ "sel_acc80": 0.9466666666666667,
1355
+ "conf_wrong": 0.0,
1356
+ "coverage": 0.9786666666666667,
1357
+ "false_abstain": 0.0,
1358
+ "null_recall": 0.9696969696969697,
1359
+ "auroc_null": 0.9999793435376257,
1360
+ "auroc_conf": 1.0
1361
+ },
1362
+ "scaled": {
1363
+ "nll": 0.349651697481338,
1364
+ "brier": 0.19284379766016657,
1365
+ "jsd": 0.10093981758344396,
1366
+ "acc": 0.878,
1367
+ "acc_k": 0.8759372869802318,
1368
+ "ece": 0.07291284207503003,
1369
+ "sel_acc80": 0.9433333333333334,
1370
+ "conf_wrong": 0.0,
1371
+ "coverage": 0.9786666666666667,
1372
+ "false_abstain": 0.0,
1373
+ "null_recall": 0.9696969696969697,
1374
+ "auroc_null": 0.9999793435376257,
1375
+ "auroc_conf": 1.0
1376
+ },
1377
+ "ordinal": {
1378
+ "ordinal_mae": 0.26497695852534564,
1379
+ "exp_score_err": 0.43905842999686256,
1380
+ "n": 434
1381
+ },
1382
+ "noul_reversed": {
1383
+ "max_abs_diff": 0.010068416595458984,
1384
+ "mean_abs_diff": 0.00043900286989631126
1385
+ }
1386
+ },
1387
+ "wf_heldout_choice": {
1388
+ "raw": {
1389
+ "nll": 0.08894784373441748,
1390
+ "brier": 0.03804468692396801,
1391
+ "jsd": 0.017131709690636845,
1392
+ "acc": 0.978,
1393
+ "acc_k": 0.9907997169143666,
1394
+ "ece": 0.011830694854259511,
1395
+ "sel_acc80": 0.9958333333333333,
1396
+ "conf_wrong": 0.0,
1397
+ "coverage": 0.9273333333333333,
1398
+ "false_abstain": 0.01556970983722576,
1399
+ "null_recall": 1.0,
1400
+ "auroc_null": 0.9994834500654839,
1401
+ "auroc_conf": 0.9995119213217171
1402
+ },
1403
+ "scaled": {
1404
+ "nll": 0.09155318678100496,
1405
+ "brier": 0.03891874334203766,
1406
+ "jsd": 0.018270416526501614,
1407
+ "acc": 0.978,
1408
+ "acc_k": 0.9907997169143666,
1409
+ "ece": 0.012840404589970836,
1410
+ "sel_acc80": 0.995,
1411
+ "conf_wrong": 0.0,
1412
+ "coverage": 0.9273333333333333,
1413
+ "false_abstain": 0.01556970983722576,
1414
+ "null_recall": 1.0,
1415
+ "auroc_null": 0.9994834500654839,
1416
+ "auroc_conf": 0.9995119213217171
1417
+ }
1418
+ },
1419
+ "wf_heldout_noul": {
1420
+ "raw": {
1421
+ "nll": 0.5731820141841979,
1422
+ "brier": 0.30167717441890574,
1423
+ "jsd": 0.10894462183976376,
1424
+ "acc": 0.8106666666666666,
1425
+ "acc_k": 0.8106666666666666,
1426
+ "ece": 0.13052070760726933,
1427
+ "sel_acc80": 0.88,
1428
+ "coverage": 1.0,
1429
+ "false_abstain": 0.0,
1430
+ "null_recall": "n/a"
1431
+ },
1432
+ "scaled": {
1433
+ "nll": 0.5272018810924806,
1434
+ "brier": 0.2944517623199594,
1435
+ "jsd": 0.10771026191000946,
1436
+ "acc": 0.8106666666666666,
1437
+ "acc_k": 0.8106666666666666,
1438
+ "ece": 0.12216018517812091,
1439
+ "sel_acc80": 0.88,
1440
+ "coverage": 1.0,
1441
+ "false_abstain": 0.0,
1442
+ "null_recall": "n/a"
1443
+ },
1444
+ "noul_reversed": {
1445
+ "max_abs_diff": 0.0,
1446
+ "mean_abs_diff": 0.0
1447
+ }
1448
+ },
1449
+ "wf_heldout_score": {
1450
+ "raw": {
1451
+ "nll": 1.0186137959697332,
1452
+ "brier": 0.5645217285335032,
1453
+ "jsd": 0.28011783812053404,
1454
+ "acc": 0.528,
1455
+ "acc_k": 0.528,
1456
+ "ece": 0.04999171092112859,
1457
+ "sel_acc80": 0.5808333333333333,
1458
+ "coverage": 1.0,
1459
+ "false_abstain": 0.0,
1460
+ "null_recall": "n/a"
1461
+ },
1462
+ "scaled": {
1463
+ "nll": 1.0259130994214425,
1464
+ "brier": 0.5673302940420648,
1465
+ "jsd": 0.2850322863655089,
1466
+ "acc": 0.528,
1467
+ "acc_k": 0.528,
1468
+ "ece": 0.06743905701239906,
1469
+ "sel_acc80": 0.5808333333333333,
1470
+ "coverage": 1.0,
1471
+ "false_abstain": 0.0,
1472
+ "null_recall": "n/a"
1473
+ },
1474
+ "ordinal": {
1475
+ "ordinal_mae": 0.5706666666666667,
1476
+ "exp_score_err": 0.6252469550896435,
1477
+ "n": 1500
1478
+ }
1479
+ },
1480
+ "wf_heldout_style": {
1481
+ "raw": {
1482
+ "nll": 0.3463679467600616,
1483
+ "brier": 0.2021043802230954,
1484
+ "jsd": 0.10170733756745631,
1485
+ "acc": 0.8606666666666667,
1486
+ "acc_k": 0.8584969532836831,
1487
+ "ece": 0.05806009686986599,
1488
+ "sel_acc80": 0.9341666666666667,
1489
+ "conf_wrong": 0.0,
1490
+ "coverage": 0.9846666666666667,
1491
+ "false_abstain": 0.0,
1492
+ "null_recall": 1.0,
1493
+ "auroc_null": 1.0,
1494
+ "auroc_conf": 1.0
1495
+ },
1496
+ "scaled": {
1497
+ "nll": 0.3576204161242343,
1498
+ "brier": 0.20659511693859356,
1499
+ "jsd": 0.10588178815700212,
1500
+ "acc": 0.8606666666666667,
1501
+ "acc_k": 0.8584969532836831,
1502
+ "ece": 0.07112260150909422,
1503
+ "sel_acc80": 0.9325,
1504
+ "conf_wrong": 0.0,
1505
+ "coverage": 0.9846666666666667,
1506
+ "false_abstain": 0.0,
1507
+ "null_recall": 1.0,
1508
+ "auroc_null": 1.0,
1509
+ "auroc_conf": 1.0
1510
+ },
1511
+ "ordinal": {
1512
+ "ordinal_mae": 0.3010752688172043,
1513
+ "exp_score_err": 0.46150196176081854,
1514
+ "n": 465
1515
+ },
1516
+ "noul_reversed": {
1517
+ "max_abs_diff": 0.012003302574157715,
1518
+ "mean_abs_diff": 0.0005715727954548829
1519
+ }
1520
+ },
1521
+ "wf_rubric_flip": {
1522
+ "raw": {
1523
+ "nll": 0.5047635419647872,
1524
+ "brier": 0.2691039167691245,
1525
+ "jsd": 0.12483415140150374,
1526
+ "acc": 0.794,
1527
+ "acc_k": 0.7986666666666666,
1528
+ "ece": 0.03788818931579589,
1529
+ "sel_acc80": 0.895,
1530
+ "coverage": 0.994,
1531
+ "false_abstain": 0.006,
1532
+ "null_recall": "n/a"
1533
+ },
1534
+ "scaled": {
1535
+ "nll": 0.49939328249927545,
1536
+ "brier": 0.26871069651279345,
1537
+ "jsd": 0.12677251415967972,
1538
+ "acc": 0.794,
1539
+ "acc_k": 0.7986666666666666,
1540
+ "ece": 0.03212277259429295,
1541
+ "sel_acc80": 0.8958333333333334,
1542
+ "coverage": 0.994,
1543
+ "false_abstain": 0.006,
1544
+ "null_recall": "n/a"
1545
+ },
1546
+ "flip": {
1547
+ "n_pairs": 750,
1548
+ "flip_rate": 0.8173333333333334,
1549
+ "both_correct": 0.6733333333333333,
1550
+ "acc": 0.7986666666666666
1551
+ },
1552
+ "ordinal": {
1553
+ "ordinal_mae": 0.5816733067729084,
1554
+ "exp_score_err": 0.6296289763572166,
1555
+ "n": 502
1556
+ },
1557
+ "noul_reversed": {
1558
+ "max_abs_diff": 0.015544265508651733,
1559
+ "mean_abs_diff": 0.0003942134614645715
1560
+ }
1561
+ },
1562
+ "wf_rubric_shuffled": {
1563
+ "raw": {
1564
+ "nll": 3.2318217026493268,
1565
+ "brier": 0.9866861870364446,
1566
+ "jsd": 0.38150478086393286,
1567
+ "acc": 0.38066666666666665,
1568
+ "acc_k": 0.382,
1569
+ "ece": 0.40178403099377946,
1570
+ "sel_acc80": 0.395,
1571
+ "coverage": 0.9906666666666667,
1572
+ "false_abstain": 0.009333333333333334,
1573
+ "null_recall": "n/a"
1574
+ },
1575
+ "scaled": {
1576
+ "nll": 2.936559013634907,
1577
+ "brier": 0.9733415497626388,
1578
+ "jsd": 0.3787540582851265,
1579
+ "acc": 0.38066666666666665,
1580
+ "acc_k": 0.382,
1581
+ "ece": 0.3852959664066633,
1582
+ "sel_acc80": 0.39666666666666667,
1583
+ "coverage": 0.9906666666666667,
1584
+ "false_abstain": 0.009333333333333334,
1585
+ "null_recall": "n/a"
1586
+ },
1587
+ "ordinal": {
1588
+ "ordinal_mae": 0.890652557319224,
1589
+ "exp_score_err": 0.8865040883733496,
1590
+ "n": 567
1591
+ },
1592
+ "noul_reversed": {
1593
+ "max_abs_diff": 0.008019983768463135,
1594
+ "mean_abs_diff": 0.00052099735086022
1595
+ }
1596
+ },
1597
+ "data_wf_hf_val": {
1598
+ "raw": {
1599
+ "nll": 0.014253304002249504,
1600
+ "brier": 0.008345681951172091,
1601
+ "jsd": 0.003829924295894302,
1602
+ "acc": 0.9946666666666667,
1603
+ "acc_k": 0.9946666666666667,
1604
+ "ece": 0.0024553716182709445,
1605
+ "sel_acc80": 1.0,
1606
+ "coverage": 1.0,
1607
+ "false_abstain": 0.0,
1608
+ "null_recall": "n/a"
1609
+ },
1610
+ "scaled": {
1611
+ "nll": 0.014647498088970042,
1612
+ "brier": 0.008344898141976866,
1613
+ "jsd": 0.004117263061674672,
1614
+ "acc": 0.9946666666666667,
1615
+ "acc_k": 0.9946666666666667,
1616
+ "ece": 0.0016058325767517242,
1617
+ "sel_acc80": 1.0,
1618
+ "coverage": 1.0,
1619
+ "false_abstain": 0.0,
1620
+ "null_recall": "n/a"
1621
+ }
1622
+ },
1623
+ "cua_s1_forms_demo": {
1624
+ "raw": {
1625
+ "nll": 0.0006697119946596961,
1626
+ "brier": 2.5804204528327477e-06,
1627
+ "jsd": 0.00023202367765799874,
1628
+ "acc": 1.0,
1629
+ "acc_k": 1.0,
1630
+ "ece": 0.0006688650773496452,
1631
+ "sel_acc80": 1.0,
1632
+ "coverage": 1.0,
1633
+ "false_abstain": 0.0,
1634
+ "null_recall": "n/a"
1635
+ },
1636
+ "scaled": {
1637
+ "nll": 0.0016030520060776182,
1638
+ "brier": 1.0894076450595428e-05,
1639
+ "jsd": 0.0005552100790503665,
1640
+ "acc": 1.0,
1641
+ "acc_k": 1.0,
1642
+ "ece": 0.0015992209011194136,
1643
+ "sel_acc80": 1.0,
1644
+ "coverage": 1.0,
1645
+ "false_abstain": 0.0,
1646
+ "null_recall": "n/a"
1647
+ }
1648
+ },
1649
+ "cua_s1_forms_test": {
1650
+ "raw": {
1651
+ "nll": 0.03166430327218538,
1652
+ "brier": 0.015101025211502896,
1653
+ "jsd": 0.0058346216304238216,
1654
+ "acc": 0.9913333333333333,
1655
+ "acc_k": 0.9913333333333333,
1656
+ "ece": 0.0068630737860996775,
1657
+ "sel_acc80": 1.0,
1658
+ "coverage": 1.0,
1659
+ "false_abstain": 0.0,
1660
+ "null_recall": "n/a"
1661
+ },
1662
+ "scaled": {
1663
+ "nll": 0.029713019758039725,
1664
+ "brier": 0.014819644050189798,
1665
+ "jsd": 0.005967063943986709,
1666
+ "acc": 0.9913333333333333,
1667
+ "acc_k": 0.9913333333333333,
1668
+ "ece": 0.005603543957074477,
1669
+ "sel_acc80": 1.0,
1670
+ "coverage": 1.0,
1671
+ "false_abstain": 0.0,
1672
+ "null_recall": "n/a"
1673
+ }
1674
+ },
1675
+ "jev4b_adversarial": {
1676
+ "raw": {
1677
+ "nll": 0.38843911469630704,
1678
+ "brier": 0.21720070872016706,
1679
+ "jsd": 0.09923600700781257,
1680
+ "acc": 0.835,
1681
+ "acc_k": 0.835,
1682
+ "ece": 0.04567231977979346,
1683
+ "sel_acc80": 0.9354166666666667,
1684
+ "coverage": 1.0,
1685
+ "false_abstain": 0.0,
1686
+ "null_recall": "n/a"
1687
+ },
1688
+ "scaled": {
1689
+ "nll": 0.3825427230468783,
1690
+ "brier": 0.21588778092143868,
1691
+ "jsd": 0.10090386299009478,
1692
+ "acc": 0.835,
1693
+ "acc_k": 0.835,
1694
+ "ece": 0.03748601749539371,
1695
+ "sel_acc80": 0.9333333333333333,
1696
+ "coverage": 1.0,
1697
+ "false_abstain": 0.0,
1698
+ "null_recall": "n/a"
1699
+ },
1700
+ "ordinal": {
1701
+ "ordinal_mae": 0.43333333333333335,
1702
+ "exp_score_err": 0.4572230171763431,
1703
+ "n": 120
1704
+ },
1705
+ "noul_reversed": {
1706
+ "max_abs_diff": 0.0018151402473449707,
1707
+ "mean_abs_diff": 0.0003225292331383874
1708
+ }
1709
+ },
1710
+ "jevlogs_triage": {
1711
+ "raw": {
1712
+ "nll": 1.2262092962906508,
1713
+ "brier": 0.7776982801860635,
1714
+ "jsd": 0.2786176452889147,
1715
+ "acc": 0.48,
1716
+ "acc_k": 0.48,
1717
+ "ece": 0.3425089488426844,
1718
+ "sel_acc80": 0.4725,
1719
+ "coverage": 1.0,
1720
+ "false_abstain": 0.0,
1721
+ "null_recall": "n/a"
1722
+ },
1723
+ "scaled": {
1724
+ "nll": 1.1367764885369938,
1725
+ "brier": 0.7531789068144168,
1726
+ "jsd": 0.2718081592157289,
1727
+ "acc": 0.48,
1728
+ "acc_k": 0.48,
1729
+ "ece": 0.3260512084364891,
1730
+ "sel_acc80": 0.4725,
1731
+ "coverage": 1.0,
1732
+ "false_abstain": 0.0,
1733
+ "null_recall": "n/a"
1734
+ },
1735
+ "noul_reversed": {
1736
+ "max_abs_diff": 0.0,
1737
+ "mean_abs_diff": 0.0
1738
+ }
1739
+ },
1740
+ "mind2web_choice": {
1741
+ "raw": {
1742
+ "nll": 1.2412872188409299,
1743
+ "brier": 0.572483286457417,
1744
+ "jsd": 0.27838292216102917,
1745
+ "acc": 0.5466666666666666,
1746
+ "acc_k": 0.5466666666666666,
1747
+ "ece": 0.0791386130352815,
1748
+ "sel_acc80": 0.6166666666666667,
1749
+ "coverage": 1.0,
1750
+ "false_abstain": 0.0,
1751
+ "null_recall": "n/a"
1752
+ },
1753
+ "scaled": {
1754
+ "nll": 1.2202121294813058,
1755
+ "brier": 0.5664862341494601,
1756
+ "jsd": 0.28233206877326084,
1757
+ "acc": 0.5466666666666666,
1758
+ "acc_k": 0.5466666666666666,
1759
+ "ece": 0.060432610829671184,
1760
+ "sel_acc80": 0.6166666666666667,
1761
+ "coverage": 1.0,
1762
+ "false_abstain": 0.0,
1763
+ "null_recall": "n/a"
1764
+ },
1765
+ "noul_reversed": {
1766
+ "max_abs_diff": 0.016718044877052307,
1767
+ "mean_abs_diff": 0.0021705068728576104
1768
+ }
1769
+ },
1770
+ "pagerduty_trigger": {
1771
+ "raw": {
1772
+ "nll": 0.8354194549929469,
1773
+ "brier": 0.5352882009354301,
1774
+ "jsd": 0.22011285498403638,
1775
+ "acc": 0.5533333333333333,
1776
+ "acc_k": 0.5533333333333333,
1777
+ "ece": 0.2813940734863281,
1778
+ "sel_acc80": 0.6325,
1779
+ "coverage": 1.0,
1780
+ "false_abstain": 0.0,
1781
+ "null_recall": "n/a"
1782
+ },
1783
+ "scaled": {
1784
+ "nll": 0.8100947578470528,
1785
+ "brier": 0.5245902994052029,
1786
+ "jsd": 0.21967471371958058,
1787
+ "acc": 0.5533333333333333,
1788
+ "acc_k": 0.5533333333333333,
1789
+ "ece": 0.24505597927172978,
1790
+ "sel_acc80": 0.6316666666666667,
1791
+ "coverage": 1.0,
1792
+ "false_abstain": 0.0,
1793
+ "null_recall": "n/a"
1794
+ },
1795
+ "noul_reversed": {
1796
+ "max_abs_diff": 0.0083027184009552,
1797
+ "mean_abs_diff": 0.0007018453646451234
1798
+ }
1799
+ },
1800
+ "systemone_lite_hard": {
1801
+ "raw": {
1802
+ "nll": 0.44872061751977815,
1803
+ "brier": 0.2415014631857414,
1804
+ "jsd": 0.1130754906430559,
1805
+ "acc": 0.8526666666666667,
1806
+ "acc_k": 0.8526666666666667,
1807
+ "ece": 0.06893132446209589,
1808
+ "sel_acc80": 0.8908333333333334,
1809
+ "coverage": 1.0,
1810
+ "false_abstain": 0.0,
1811
+ "null_recall": "n/a"
1812
+ },
1813
+ "scaled": {
1814
+ "nll": 0.44582079406698827,
1815
+ "brier": 0.24517414554091083,
1816
+ "jsd": 0.11670442592270194,
1817
+ "acc": 0.8526666666666667,
1818
+ "acc_k": 0.8526666666666667,
1819
+ "ece": 0.07723588303724928,
1820
+ "sel_acc80": 0.8908333333333334,
1821
+ "coverage": 1.0,
1822
+ "false_abstain": 0.0,
1823
+ "null_recall": "n/a"
1824
+ },
1825
+ "ordinal": {
1826
+ "ordinal_mae": 0.182,
1827
+ "exp_score_err": 0.35481958847247086,
1828
+ "n": 500
1829
+ },
1830
+ "noul_reversed": {
1831
+ "max_abs_diff": 0.014602690935134888,
1832
+ "mean_abs_diff": 0.001006261961236305
1833
+ }
1834
+ },
1835
+ "tree_choice_cap": {
1836
+ "raw": {
1837
+ "nll": 1.0944042301182224,
1838
+ "brier": 0.41195583724182677,
1839
+ "jsd": 0.22831650697305347,
1840
+ "acc": 0.6902777777777778,
1841
+ "acc_k": 0.7319444444444444,
1842
+ "ece": 0.07779081524349749,
1843
+ "sel_acc80": 0.7934027777777778,
1844
+ "coverage": 0.8611111111111112,
1845
+ "false_abstain": 0.1388888888888889,
1846
+ "null_recall": "n/a"
1847
+ },
1848
+ "scaled": {
1849
+ "nll": 1.1347869650083346,
1850
+ "brier": 0.4228106671667945,
1851
+ "jsd": 0.23927792250464092,
1852
+ "acc": 0.6833333333333333,
1853
+ "acc_k": 0.7319444444444444,
1854
+ "ece": 0.0688384915339864,
1855
+ "sel_acc80": 0.7881944444444444,
1856
+ "coverage": 0.85,
1857
+ "false_abstain": 0.15,
1858
+ "null_recall": "n/a"
1859
+ }
1860
+ },
1861
+ "typed_decisions_test": {
1862
+ "raw": {
1863
+ "nll": 1.2050029006817273,
1864
+ "brier": 0.23359803910932034,
1865
+ "jsd": 0.10082391914655038,
1866
+ "acc": 0.54,
1867
+ "acc_k": 0.54,
1868
+ "ece": 0.08852838679154712,
1869
+ "sel_acc80": 0.5691666666666667,
1870
+ "coverage": 1.0,
1871
+ "false_abstain": 0.0,
1872
+ "null_recall": "n/a"
1873
+ },
1874
+ "scaled": {
1875
+ "nll": 1.1675730711137542,
1876
+ "brier": 0.22356043087364605,
1877
+ "jsd": 0.09743992727868328,
1878
+ "acc": 0.54,
1879
+ "acc_k": 0.54,
1880
+ "ece": 0.06417149796088538,
1881
+ "sel_acc80": 0.5683333333333334,
1882
+ "coverage": 1.0,
1883
+ "false_abstain": 0.0,
1884
+ "null_recall": "n/a"
1885
+ },
1886
+ "ordinal": {
1887
+ "ordinal_mae": 0.5883333333333334,
1888
+ "exp_score_err": 0.7085446309794221,
1889
+ "n": 600
1890
+ },
1891
+ "noul_reversed": {
1892
+ "max_abs_diff": 0.013534486293792725,
1893
+ "mean_abs_diff": 0.0021557581052184106
1894
+ }
1895
+ },
1896
+ "typed_decisions_train": {
1897
+ "raw": {
1898
+ "nll": 1.416492816208917,
1899
+ "brier": 0.25422257216150745,
1900
+ "jsd": 0.10777533042065943,
1901
+ "acc": 0.4266666666666667,
1902
+ "acc_k": 0.4266666666666667,
1903
+ "ece": 0.18357078438997268,
1904
+ "sel_acc80": 0.42916666666666664,
1905
+ "coverage": 1.0,
1906
+ "false_abstain": 0.0,
1907
+ "null_recall": "n/a"
1908
+ },
1909
+ "scaled": {
1910
+ "nll": 1.3595566999012847,
1911
+ "brier": 0.2359260337969907,
1912
+ "jsd": 0.1005492616508014,
1913
+ "acc": 0.4266666666666667,
1914
+ "acc_k": 0.4266666666666667,
1915
+ "ece": 0.1572492956717809,
1916
+ "sel_acc80": 0.4266666666666667,
1917
+ "coverage": 1.0,
1918
+ "false_abstain": 0.0,
1919
+ "null_recall": "n/a"
1920
+ },
1921
+ "ordinal": {
1922
+ "ordinal_mae": 0.75,
1923
+ "exp_score_err": 0.7864355162782587,
1924
+ "n": 600
1925
+ },
1926
+ "noul_reversed": {
1927
+ "max_abs_diff": 0.011488527059555054,
1928
+ "mean_abs_diff": 0.003036180833975474
1929
+ }
1930
+ },
1931
+ "data_wh_val": {
1932
+ "raw": {
1933
+ "nll": 0.2300384575464517,
1934
+ "brier": 0.14204218823632153,
1935
+ "jsd": 0.06341173013641205,
1936
+ "acc": 0.894,
1937
+ "acc_k": 0.9278737791134485,
1938
+ "ece": 0.015060881972312945,
1939
+ "sel_acc80": 0.9716666666666667,
1940
+ "conf_wrong": 0.005917159763313609,
1941
+ "coverage": 0.8546666666666667,
1942
+ "false_abstain": 0.048835462058602556,
1943
+ "null_recall": 0.9053254437869822,
1944
+ "auroc_null": 0.9826530748336214,
1945
+ "auroc_conf": 0.9783230120165911
1946
+ },
1947
+ "scaled": {
1948
+ "nll": 0.23008159136699488,
1949
+ "brier": 0.14167703723149266,
1950
+ "jsd": 0.0641277525235366,
1951
+ "acc": 0.8926666666666667,
1952
+ "acc_k": 0.9278737791134485,
1953
+ "ece": 0.016159646193186408,
1954
+ "sel_acc80": 0.9733333333333334,
1955
+ "conf_wrong": 0.005917159763313609,
1956
+ "coverage": 0.8533333333333333,
1957
+ "false_abstain": 0.05033809166040571,
1958
+ "null_recall": 0.9053254437869822,
1959
+ "auroc_null": 0.9826530748336215,
1960
+ "auroc_conf": 0.9778339905485487
1961
+ }
1962
+ },
1963
+ "wh_heldout_family": {
1964
+ "raw": {
1965
+ "nll": 0.3386832933282492,
1966
+ "brier": 0.19001383971199456,
1967
+ "jsd": 0.08311882716530435,
1968
+ "acc": 0.8649289099526066,
1969
+ "acc_k": 0.8938992042440318,
1970
+ "ece": 0.04368148955971149,
1971
+ "sel_acc80": 0.9408284023668639,
1972
+ "conf_wrong": 0.0,
1973
+ "coverage": 0.8625592417061612,
1974
+ "false_abstain": 0.04774535809018567,
1975
+ "null_recall": 0.8888888888888888,
1976
+ "auroc_null": 0.9831417624521073,
1977
+ "auroc_conf": 0.9725316828765105
1978
+ },
1979
+ "scaled": {
1980
+ "nll": 0.3327042389265942,
1981
+ "brier": 0.1900852589651073,
1982
+ "jsd": 0.08425147199473242,
1983
+ "acc": 0.8649289099526066,
1984
+ "acc_k": 0.8938992042440318,
1985
+ "ece": 0.04666075068062519,
1986
+ "sel_acc80": 0.9408284023668639,
1987
+ "conf_wrong": 0.0,
1988
+ "coverage": 0.8625592417061612,
1989
+ "false_abstain": 0.04774535809018567,
1990
+ "null_recall": 0.8888888888888888,
1991
+ "auroc_null": 0.9831417624521073,
1992
+ "auroc_conf": 0.9718243442381373
1993
+ }
1994
+ },
1995
+ "wh_heldout_grammar": {
1996
+ "raw": {
1997
+ "nll": 0.2637673505050913,
1998
+ "brier": 0.15817524496433566,
1999
+ "jsd": 0.07089041621921395,
2000
+ "acc": 0.8799019607843137,
2001
+ "acc_k": 0.914364640883978,
2002
+ "ece": 0.024724171843891023,
2003
+ "sel_acc80": 0.9601837672281777,
2004
+ "conf_wrong": 0.0,
2005
+ "coverage": 0.8480392156862745,
2006
+ "false_abstain": 0.05386740331491713,
2007
+ "null_recall": 0.9239130434782609,
2008
+ "auroc_null": 0.9828849387460965,
2009
+ "auroc_conf": 0.9772249579630075
2010
+ },
2011
+ "scaled": {
2012
+ "nll": 0.2635058749685546,
2013
+ "brier": 0.15850269393002997,
2014
+ "jsd": 0.07194949020789403,
2015
+ "acc": 0.8799019607843137,
2016
+ "acc_k": 0.914364640883978,
2017
+ "ece": 0.025690338117819185,
2018
+ "sel_acc80": 0.9586523736600306,
2019
+ "conf_wrong": 0.0,
2020
+ "coverage": 0.8480392156862745,
2021
+ "false_abstain": 0.05386740331491713,
2022
+ "null_recall": 0.9239130434782609,
2023
+ "auroc_null": 0.9828849387460965,
2024
+ "auroc_conf": 0.9767145087677156
2025
+ }
2026
+ },
2027
+ "wh_heldout_style": {
2028
+ "raw": {
2029
+ "nll": 0.26686252182988546,
2030
+ "brier": 0.15311981455702697,
2031
+ "jsd": 0.06948256408727153,
2032
+ "acc": 0.8858560794044665,
2033
+ "acc_k": 0.8858560794044665,
2034
+ "ece": 0.03080959015389527,
2035
+ "sel_acc80": 0.9752321981424149,
2036
+ "coverage": 1.0,
2037
+ "false_abstain": 0.0,
2038
+ "null_recall": "n/a"
2039
+ },
2040
+ "scaled": {
2041
+ "nll": 0.2633042948240913,
2042
+ "brier": 0.1516918189603084,
2043
+ "jsd": 0.07049474346320594,
2044
+ "acc": 0.8858560794044665,
2045
+ "acc_k": 0.8858560794044665,
2046
+ "ece": 0.029319018938996952,
2047
+ "sel_acc80": 0.978328173374613,
2048
+ "coverage": 1.0,
2049
+ "false_abstain": 0.0,
2050
+ "null_recall": "n/a"
2051
+ }
2052
+ },
2053
+ "wh_level7": {
2054
+ "raw": {
2055
+ "nll": 0.8425315491747045,
2056
+ "brier": 0.566066637960518,
2057
+ "jsd": 0.22709053783966102,
2058
+ "acc": 0.5443181818181818,
2059
+ "acc_k": 0.5443181818181818,
2060
+ "ece": 0.17712931768460705,
2061
+ "sel_acc80": 0.5738636363636364,
2062
+ "coverage": 1.0,
2063
+ "false_abstain": 0.0,
2064
+ "null_recall": "n/a"
2065
+ },
2066
+ "scaled": {
2067
+ "nll": 0.815751176572115,
2068
+ "brier": 0.5531539212337735,
2069
+ "jsd": 0.22560885355170332,
2070
+ "acc": 0.5443181818181818,
2071
+ "acc_k": 0.5443181818181818,
2072
+ "ece": 0.15701854222200134,
2073
+ "sel_acc80": 0.5724431818181818,
2074
+ "coverage": 1.0,
2075
+ "false_abstain": 0.0,
2076
+ "null_recall": "n/a"
2077
+ }
2078
+ },
2079
+ "wh_rubric_flip": {
2080
+ "raw": {
2081
+ "nll": 0.37024391536252643,
2082
+ "brier": 0.2251475372121566,
2083
+ "jsd": 0.09686676026339679,
2084
+ "acc": 0.832,
2085
+ "acc_k": 0.8517715112075199,
2086
+ "ece": 0.03836428254842761,
2087
+ "sel_acc80": 0.9116666666666666,
2088
+ "conf_wrong": 0.0,
2089
+ "coverage": 0.898,
2090
+ "false_abstain": 0.03398409255242227,
2091
+ "null_recall": 0.905982905982906,
2092
+ "auroc_null": 0.9883258863736086,
2093
+ "auroc_conf": 0.9758978067004097
2094
+ },
2095
+ "scaled": {
2096
+ "nll": 0.36406225678752874,
2097
+ "brier": 0.22335213586456443,
2098
+ "jsd": 0.09762219605986137,
2099
+ "acc": 0.8326666666666667,
2100
+ "acc_k": 0.8517715112075199,
2101
+ "ece": 0.03447558822234474,
2102
+ "sel_acc80": 0.9125,
2103
+ "conf_wrong": 0.0,
2104
+ "coverage": 0.8973333333333333,
2105
+ "false_abstain": 0.03398409255242227,
2106
+ "null_recall": 0.9145299145299145,
2107
+ "auroc_null": 0.9883258863736086,
2108
+ "auroc_conf": 0.9749893394145022
2109
+ }
2110
+ },
2111
+ "wh_rubric_shuffled": {
2112
+ "raw": {
2113
+ "nll": 3.3784176615021364,
2114
+ "brier": 1.0176747777780963,
2115
+ "jsd": 0.36765612382768675,
2116
+ "acc": 0.374,
2117
+ "acc_k": 0.4036697247706422,
2118
+ "ece": 0.466156586488088,
2119
+ "sel_acc80": 0.3441666666666667,
2120
+ "conf_wrong": 0.18072289156626506,
2121
+ "coverage": 0.9299999999999999,
2122
+ "false_abstain": 0.0465772759350741,
2123
+ "null_recall": 0.46987951807228917,
2124
+ "auroc_null": 0.9583967486034469,
2125
+ "auroc_conf": 0.7938543163479606
2126
+ },
2127
+ "scaled": {
2128
+ "nll": 3.0621024358996203,
2129
+ "brier": 1.0025028615908116,
2130
+ "jsd": 0.36380861609352405,
2131
+ "acc": 0.37333333333333335,
2132
+ "acc_k": 0.4036697247706422,
2133
+ "ece": 0.4562540921767553,
2134
+ "sel_acc80": 0.3458333333333333,
2135
+ "conf_wrong": 0.1686746987951807,
2136
+ "coverage": 0.9293333333333333,
2137
+ "false_abstain": 0.04728299223712068,
2138
+ "null_recall": 0.46987951807228917,
2139
+ "auroc_null": 0.9583967486034469,
2140
+ "auroc_conf": 0.7857258249653519
2141
+ }
2142
+ },
2143
+ "data_u_val": {
2144
+ "raw": {
2145
+ "nll": 0.4886842598845047,
2146
+ "brier": 0.04234767853598601,
2147
+ "jsd": 0.018090470816900562,
2148
+ "acc": 0.8953333333333333,
2149
+ "acc_k": 0.8953333333333333,
2150
+ "ece": 0.12599515644709272,
2151
+ "sel_acc80": 0.9616666666666667,
2152
+ "coverage": 1.0,
2153
+ "false_abstain": 0.0,
2154
+ "null_recall": "n/a"
2155
+ },
2156
+ "scaled": {
2157
+ "nll": 0.49177310430162957,
2158
+ "brier": 0.04392926004492564,
2159
+ "jsd": 0.019119015669118427,
2160
+ "acc": 0.8953333333333333,
2161
+ "acc_k": 0.8953333333333333,
2162
+ "ece": 0.14489340595404307,
2163
+ "sel_acc80": 0.9616666666666667,
2164
+ "coverage": 1.0,
2165
+ "false_abstain": 0.0,
2166
+ "null_recall": "n/a"
2167
+ }
2168
+ },
2169
+ "u_chaosnli": {
2170
+ "raw": {
2171
+ "nll": 1.0139885174394003,
2172
+ "brier": 0.17014809164519165,
2173
+ "jsd": 0.06626210261670297,
2174
+ "acc": 0.5166666666666667,
2175
+ "acc_k": 0.5166666666666667,
2176
+ "ece": 0.10358940768241887,
2177
+ "sel_acc80": 0.5341666666666667,
2178
+ "coverage": 1.0,
2179
+ "false_abstain": 0.0,
2180
+ "null_recall": "n/a"
2181
+ },
2182
+ "scaled": {
2183
+ "nll": 1.0020134621764547,
2184
+ "brier": 0.16190185540335636,
2185
+ "jsd": 0.06505670021636623,
2186
+ "acc": 0.5166666666666667,
2187
+ "acc_k": 0.5166666666666667,
2188
+ "ece": 0.08189480453729626,
2189
+ "sel_acc80": 0.5341666666666667,
2190
+ "coverage": 1.0,
2191
+ "false_abstain": 0.0,
2192
+ "null_recall": "n/a"
2193
+ }
2194
+ },
2195
+ "u_real_heldout": {
2196
+ "raw": {
2197
+ "nll": 0.6395687144671027,
2198
+ "brier": 0.11684180579038546,
2199
+ "jsd": 0.055302365340723074,
2200
+ "acc": 0.6583629893238434,
2201
+ "acc_k": 0.6583629893238434,
2202
+ "ece": 0.04802694863696116,
2203
+ "sel_acc80": 0.7,
2204
+ "coverage": 1.0,
2205
+ "false_abstain": 0.0,
2206
+ "null_recall": "n/a"
2207
+ },
2208
+ "scaled": {
2209
+ "nll": 0.6408081760955402,
2210
+ "brier": 0.11675911344973736,
2211
+ "jsd": 0.05675639741980845,
2212
+ "acc": 0.6583629893238434,
2213
+ "acc_k": 0.6583629893238434,
2214
+ "ece": 0.06303854366214252,
2215
+ "sel_acc80": 0.7022222222222222,
2216
+ "coverage": 1.0,
2217
+ "false_abstain": 0.0,
2218
+ "null_recall": "n/a"
2219
+ }
2220
+ },
2221
+ "u_synthetic_heldout": {
2222
+ "raw": {
2223
+ "nll": 0.6619388393182489,
2224
+ "brier": 0.029335292468287227,
2225
+ "jsd": 0.013479520479808405,
2226
+ "acc": 0.9073684210526316,
2227
+ "acc_k": 0.9073684210526316,
2228
+ "ece": 0.17639653883482279,
2229
+ "sel_acc80": 0.9815789473684211,
2230
+ "coverage": 1.0,
2231
+ "false_abstain": 0.0,
2232
+ "null_recall": "n/a"
2233
+ },
2234
+ "scaled": {
2235
+ "nll": 0.6541064203422494,
2236
+ "brier": 0.025880943825774404,
2237
+ "jsd": 0.012156440703303576,
2238
+ "acc": 0.9073684210526316,
2239
+ "acc_k": 0.9073684210526316,
2240
+ "ece": 0.19372765189722962,
2241
+ "sel_acc80": 0.9789473684210527,
2242
+ "coverage": 1.0,
2243
+ "false_abstain": 0.0,
2244
+ "null_recall": "n/a"
2245
+ }
2246
+ }
2247
+ }
2248
+ }