@mailwoman/neural-weights-en-us 5.3.0 → 5.4.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "model": "neural-weights-en-us",
3
- "model_version": "4.0.0",
3
+ "model_version": "5.3.0",
4
4
  "method": "per-locale isotonic (PAVA)",
5
5
  "tables": {
6
6
  "DE": [
@@ -8,121 +8,121 @@
8
8
  "lo": 0.0,
9
9
  "hi": 0.05,
10
10
  "center": 0.025,
11
- "calibrated": 0.8333333333333334
11
+ "calibrated": 0.860969387755102
12
12
  },
13
13
  {
14
14
  "lo": 0.05,
15
15
  "hi": 0.1,
16
16
  "center": 0.07500000000000001,
17
- "calibrated": 0.8333333333333334
17
+ "calibrated": 0.860969387755102
18
18
  },
19
19
  {
20
20
  "lo": 0.1,
21
21
  "hi": 0.15000000000000002,
22
22
  "center": 0.125,
23
- "calibrated": 0.8333333333333334
23
+ "calibrated": 0.860969387755102
24
24
  },
25
25
  {
26
26
  "lo": 0.15000000000000002,
27
27
  "hi": 0.2,
28
28
  "center": 0.17500000000000002,
29
- "calibrated": 0.8333333333333334
29
+ "calibrated": 0.860969387755102
30
30
  },
31
31
  {
32
32
  "lo": 0.2,
33
33
  "hi": 0.25,
34
34
  "center": 0.225,
35
- "calibrated": 0.8333333333333334
35
+ "calibrated": 0.860969387755102
36
36
  },
37
37
  {
38
38
  "lo": 0.25,
39
39
  "hi": 0.30000000000000004,
40
40
  "center": 0.275,
41
- "calibrated": 0.8333333333333334
41
+ "calibrated": 0.860969387755102
42
42
  },
43
43
  {
44
44
  "lo": 0.30000000000000004,
45
45
  "hi": 0.35000000000000003,
46
46
  "center": 0.32500000000000007,
47
- "calibrated": 0.8333333333333334
47
+ "calibrated": 0.860969387755102
48
48
  },
49
49
  {
50
50
  "lo": 0.35000000000000003,
51
51
  "hi": 0.4,
52
52
  "center": 0.375,
53
- "calibrated": 0.8333333333333334
53
+ "calibrated": 0.860969387755102
54
54
  },
55
55
  {
56
56
  "lo": 0.4,
57
57
  "hi": 0.45,
58
58
  "center": 0.42500000000000004,
59
- "calibrated": 0.8333333333333334
59
+ "calibrated": 0.860969387755102
60
60
  },
61
61
  {
62
62
  "lo": 0.45,
63
63
  "hi": 0.5,
64
64
  "center": 0.475,
65
- "calibrated": 0.8333333333333334
65
+ "calibrated": 0.860969387755102
66
66
  },
67
67
  {
68
68
  "lo": 0.5,
69
69
  "hi": 0.55,
70
70
  "center": 0.525,
71
- "calibrated": 0.8333333333333334
71
+ "calibrated": 0.860969387755102
72
72
  },
73
73
  {
74
74
  "lo": 0.55,
75
75
  "hi": 0.6000000000000001,
76
76
  "center": 0.5750000000000001,
77
- "calibrated": 0.8333333333333334
77
+ "calibrated": 0.860969387755102
78
78
  },
79
79
  {
80
80
  "lo": 0.6000000000000001,
81
81
  "hi": 0.65,
82
82
  "center": 0.625,
83
- "calibrated": 0.8333333333333334
83
+ "calibrated": 0.860969387755102
84
84
  },
85
85
  {
86
86
  "lo": 0.65,
87
87
  "hi": 0.7000000000000001,
88
88
  "center": 0.675,
89
- "calibrated": 0.8333333333333334
89
+ "calibrated": 0.860969387755102
90
90
  },
91
91
  {
92
92
  "lo": 0.7000000000000001,
93
93
  "hi": 0.75,
94
94
  "center": 0.7250000000000001,
95
- "calibrated": 0.8524173027989822
95
+ "calibrated": 0.860969387755102
96
96
  },
97
97
  {
98
98
  "lo": 0.75,
99
99
  "hi": 0.8,
100
100
  "center": 0.775,
101
- "calibrated": 0.8524173027989822
101
+ "calibrated": 0.860969387755102
102
102
  },
103
103
  {
104
104
  "lo": 0.8,
105
105
  "hi": 0.8500000000000001,
106
106
  "center": 0.8250000000000001,
107
- "calibrated": 0.8524173027989822
107
+ "calibrated": 0.860969387755102
108
108
  },
109
109
  {
110
110
  "lo": 0.8500000000000001,
111
111
  "hi": 0.9,
112
112
  "center": 0.875,
113
- "calibrated": 0.8524173027989822
113
+ "calibrated": 0.860969387755102
114
114
  },
115
115
  {
116
116
  "lo": 0.9,
117
117
  "hi": 0.9500000000000001,
118
118
  "center": 0.925,
119
- "calibrated": 0.8524173027989822
119
+ "calibrated": 0.860969387755102
120
120
  },
121
121
  {
122
122
  "lo": 0.9500000000000001,
123
123
  "hi": 1.0,
124
124
  "center": 0.9750000000000001,
125
- "calibrated": 0.8524173027989822
125
+ "calibrated": 0.860969387755102
126
126
  }
127
127
  ],
128
128
  "FR": [
@@ -130,121 +130,121 @@
130
130
  "lo": 0.0,
131
131
  "hi": 0.05,
132
132
  "center": 0.025,
133
- "calibrated": 0.7916666666666666
133
+ "calibrated": 0.0
134
134
  },
135
135
  {
136
136
  "lo": 0.05,
137
137
  "hi": 0.1,
138
138
  "center": 0.07500000000000001,
139
- "calibrated": 0.7916666666666666
139
+ "calibrated": 0.0906967895543893
140
140
  },
141
141
  {
142
142
  "lo": 0.1,
143
143
  "hi": 0.15000000000000002,
144
144
  "center": 0.125,
145
- "calibrated": 0.7916666666666666
145
+ "calibrated": 0.1910171464463449
146
146
  },
147
147
  {
148
148
  "lo": 0.15000000000000002,
149
149
  "hi": 0.2,
150
150
  "center": 0.17500000000000002,
151
- "calibrated": 0.7916666666666666
151
+ "calibrated": 0.29133750333830055
152
152
  },
153
153
  {
154
154
  "lo": 0.2,
155
155
  "hi": 0.25,
156
156
  "center": 0.225,
157
- "calibrated": 0.7916666666666666
157
+ "calibrated": 0.3333333333333333
158
158
  },
159
159
  {
160
160
  "lo": 0.25,
161
161
  "hi": 0.30000000000000004,
162
162
  "center": 0.275,
163
- "calibrated": 0.7916666666666666
163
+ "calibrated": 0.3333333333333333
164
164
  },
165
165
  {
166
166
  "lo": 0.30000000000000004,
167
167
  "hi": 0.35000000000000003,
168
168
  "center": 0.32500000000000007,
169
- "calibrated": 0.7916666666666666
169
+ "calibrated": 0.3333333333333333
170
170
  },
171
171
  {
172
172
  "lo": 0.35000000000000003,
173
173
  "hi": 0.4,
174
174
  "center": 0.375,
175
- "calibrated": 0.7916666666666666
175
+ "calibrated": 0.3333333333333333
176
176
  },
177
177
  {
178
178
  "lo": 0.4,
179
179
  "hi": 0.45,
180
180
  "center": 0.42500000000000004,
181
- "calibrated": 0.7916666666666666
181
+ "calibrated": 0.3333333333333333
182
182
  },
183
183
  {
184
184
  "lo": 0.45,
185
185
  "hi": 0.5,
186
186
  "center": 0.475,
187
- "calibrated": 0.7916666666666666
187
+ "calibrated": 0.7142857142857143
188
188
  },
189
189
  {
190
190
  "lo": 0.5,
191
191
  "hi": 0.55,
192
192
  "center": 0.525,
193
- "calibrated": 0.9
193
+ "calibrated": 0.7142857142857143
194
194
  },
195
195
  {
196
196
  "lo": 0.55,
197
197
  "hi": 0.6000000000000001,
198
198
  "center": 0.5750000000000001,
199
- "calibrated": 0.9508196721311475
199
+ "calibrated": 0.8444444444444444
200
200
  },
201
201
  {
202
202
  "lo": 0.6000000000000001,
203
203
  "hi": 0.65,
204
204
  "center": 0.625,
205
- "calibrated": 0.9508196721311475
205
+ "calibrated": 0.8444444444444444
206
206
  },
207
207
  {
208
208
  "lo": 0.65,
209
209
  "hi": 0.7000000000000001,
210
210
  "center": 0.675,
211
- "calibrated": 0.9508196721311475
211
+ "calibrated": 0.963963963963964
212
212
  },
213
213
  {
214
214
  "lo": 0.7000000000000001,
215
215
  "hi": 0.75,
216
216
  "center": 0.7250000000000001,
217
- "calibrated": 0.9523809523809523
217
+ "calibrated": 0.963963963963964
218
218
  },
219
219
  {
220
220
  "lo": 0.75,
221
221
  "hi": 0.8,
222
222
  "center": 0.775,
223
- "calibrated": 0.9523809523809523
223
+ "calibrated": 0.963963963963964
224
224
  },
225
225
  {
226
226
  "lo": 0.8,
227
227
  "hi": 0.8500000000000001,
228
228
  "center": 0.8250000000000001,
229
- "calibrated": 0.9523809523809523
229
+ "calibrated": 0.9736842105263158
230
230
  },
231
231
  {
232
232
  "lo": 0.8500000000000001,
233
233
  "hi": 0.9,
234
234
  "center": 0.875,
235
- "calibrated": 0.9523809523809523
235
+ "calibrated": 0.9868421052631579
236
236
  },
237
237
  {
238
238
  "lo": 0.9,
239
239
  "hi": 0.9500000000000001,
240
240
  "center": 0.925,
241
- "calibrated": 0.9843614248479583
241
+ "calibrated": 0.9891258887494772
242
242
  },
243
243
  {
244
244
  "lo": 0.9500000000000001,
245
245
  "hi": 1.0,
246
246
  "center": 0.9750000000000001,
247
- "calibrated": 0.9916666666666667
247
+ "calibrated": 1.0
248
248
  }
249
249
  ],
250
250
  "NL": [
@@ -252,115 +252,115 @@
252
252
  "lo": 0.0,
253
253
  "hi": 0.05,
254
254
  "center": 0.025,
255
- "calibrated": 0.2
255
+ "calibrated": 1.0
256
256
  },
257
257
  {
258
258
  "lo": 0.05,
259
259
  "hi": 0.1,
260
260
  "center": 0.07500000000000001,
261
- "calibrated": 0.2
261
+ "calibrated": 1.0
262
262
  },
263
263
  {
264
264
  "lo": 0.1,
265
265
  "hi": 0.15000000000000002,
266
266
  "center": 0.125,
267
- "calibrated": 0.2
267
+ "calibrated": 1.0
268
268
  },
269
269
  {
270
270
  "lo": 0.15000000000000002,
271
271
  "hi": 0.2,
272
272
  "center": 0.17500000000000002,
273
- "calibrated": 0.2
273
+ "calibrated": 1.0
274
274
  },
275
275
  {
276
276
  "lo": 0.2,
277
277
  "hi": 0.25,
278
278
  "center": 0.225,
279
- "calibrated": 0.2
279
+ "calibrated": 1.0
280
280
  },
281
281
  {
282
282
  "lo": 0.25,
283
283
  "hi": 0.30000000000000004,
284
284
  "center": 0.275,
285
- "calibrated": 0.2
285
+ "calibrated": 1.0
286
286
  },
287
287
  {
288
288
  "lo": 0.30000000000000004,
289
289
  "hi": 0.35000000000000003,
290
290
  "center": 0.32500000000000007,
291
- "calibrated": 0.2
291
+ "calibrated": 1.0
292
292
  },
293
293
  {
294
294
  "lo": 0.35000000000000003,
295
295
  "hi": 0.4,
296
296
  "center": 0.375,
297
- "calibrated": 0.8947368421052632
297
+ "calibrated": 1.0
298
298
  },
299
299
  {
300
300
  "lo": 0.4,
301
301
  "hi": 0.45,
302
302
  "center": 0.42500000000000004,
303
- "calibrated": 0.8947368421052632
303
+ "calibrated": 1.0
304
304
  },
305
305
  {
306
306
  "lo": 0.45,
307
307
  "hi": 0.5,
308
308
  "center": 0.475,
309
- "calibrated": 0.8947368421052632
309
+ "calibrated": 1.0
310
310
  },
311
311
  {
312
312
  "lo": 0.5,
313
313
  "hi": 0.55,
314
314
  "center": 0.525,
315
- "calibrated": 0.8947368421052632
315
+ "calibrated": 1.0
316
316
  },
317
317
  {
318
318
  "lo": 0.55,
319
319
  "hi": 0.6000000000000001,
320
320
  "center": 0.5750000000000001,
321
- "calibrated": 0.9985590778097982
321
+ "calibrated": 1.0
322
322
  },
323
323
  {
324
324
  "lo": 0.6000000000000001,
325
325
  "hi": 0.65,
326
326
  "center": 0.625,
327
- "calibrated": 0.9985590778097982
327
+ "calibrated": 1.0
328
328
  },
329
329
  {
330
330
  "lo": 0.65,
331
331
  "hi": 0.7000000000000001,
332
332
  "center": 0.675,
333
- "calibrated": 0.9985590778097982
333
+ "calibrated": 1.0
334
334
  },
335
335
  {
336
336
  "lo": 0.7000000000000001,
337
337
  "hi": 0.75,
338
338
  "center": 0.7250000000000001,
339
- "calibrated": 0.9985590778097982
339
+ "calibrated": 1.0
340
340
  },
341
341
  {
342
342
  "lo": 0.75,
343
343
  "hi": 0.8,
344
344
  "center": 0.775,
345
- "calibrated": 0.9985590778097982
345
+ "calibrated": 1.0
346
346
  },
347
347
  {
348
348
  "lo": 0.8,
349
349
  "hi": 0.8500000000000001,
350
350
  "center": 0.8250000000000001,
351
- "calibrated": 0.9985590778097982
351
+ "calibrated": 1.0
352
352
  },
353
353
  {
354
354
  "lo": 0.8500000000000001,
355
355
  "hi": 0.9,
356
356
  "center": 0.875,
357
- "calibrated": 0.9985590778097982
357
+ "calibrated": 1.0
358
358
  },
359
359
  {
360
360
  "lo": 0.9,
361
361
  "hi": 0.9500000000000001,
362
362
  "center": 0.925,
363
- "calibrated": 0.9985590778097982
363
+ "calibrated": 1.0
364
364
  },
365
365
  {
366
366
  "lo": 0.9500000000000001,
@@ -374,115 +374,115 @@
374
374
  "lo": 0.0,
375
375
  "hi": 0.05,
376
376
  "center": 0.025,
377
- "calibrated": 0.0
377
+ "calibrated": 0.5714285714285714
378
378
  },
379
379
  {
380
380
  "lo": 0.05,
381
381
  "hi": 0.1,
382
382
  "center": 0.07500000000000001,
383
- "calibrated": 0.0
383
+ "calibrated": 0.5714285714285714
384
384
  },
385
385
  {
386
386
  "lo": 0.1,
387
387
  "hi": 0.15000000000000002,
388
388
  "center": 0.125,
389
- "calibrated": 0.0
389
+ "calibrated": 0.5714285714285714
390
390
  },
391
391
  {
392
392
  "lo": 0.15000000000000002,
393
393
  "hi": 0.2,
394
394
  "center": 0.17500000000000002,
395
- "calibrated": 0.3333333333333333
395
+ "calibrated": 0.5714285714285714
396
396
  },
397
397
  {
398
398
  "lo": 0.2,
399
399
  "hi": 0.25,
400
400
  "center": 0.225,
401
- "calibrated": 0.3333333333333333
401
+ "calibrated": 0.6666666666666666
402
402
  },
403
403
  {
404
404
  "lo": 0.25,
405
405
  "hi": 0.30000000000000004,
406
406
  "center": 0.275,
407
- "calibrated": 0.3333333333333333
407
+ "calibrated": 0.6666666666666666
408
408
  },
409
409
  {
410
410
  "lo": 0.30000000000000004,
411
411
  "hi": 0.35000000000000003,
412
412
  "center": 0.32500000000000007,
413
- "calibrated": 0.417910447761194
413
+ "calibrated": 0.6721311475409836
414
414
  },
415
415
  {
416
416
  "lo": 0.35000000000000003,
417
417
  "hi": 0.4,
418
418
  "center": 0.375,
419
- "calibrated": 0.417910447761194
419
+ "calibrated": 0.6721311475409836
420
420
  },
421
421
  {
422
422
  "lo": 0.4,
423
423
  "hi": 0.45,
424
424
  "center": 0.42500000000000004,
425
- "calibrated": 0.48717948717948717
425
+ "calibrated": 0.6721311475409836
426
426
  },
427
427
  {
428
428
  "lo": 0.45,
429
429
  "hi": 0.5,
430
430
  "center": 0.475,
431
- "calibrated": 0.6785714285714286
431
+ "calibrated": 0.8850574712643678
432
432
  },
433
433
  {
434
434
  "lo": 0.5,
435
435
  "hi": 0.55,
436
436
  "center": 0.525,
437
- "calibrated": 0.75
437
+ "calibrated": 0.8850574712643678
438
438
  },
439
439
  {
440
440
  "lo": 0.55,
441
441
  "hi": 0.6000000000000001,
442
442
  "center": 0.5750000000000001,
443
- "calibrated": 0.7755102040816326
443
+ "calibrated": 0.8850574712643678
444
444
  },
445
445
  {
446
446
  "lo": 0.6000000000000001,
447
447
  "hi": 0.65,
448
448
  "center": 0.625,
449
- "calibrated": 0.8554913294797688
449
+ "calibrated": 0.8850574712643678
450
450
  },
451
451
  {
452
452
  "lo": 0.65,
453
453
  "hi": 0.7000000000000001,
454
454
  "center": 0.675,
455
- "calibrated": 0.8984547461368654
455
+ "calibrated": 0.8850574712643678
456
456
  },
457
457
  {
458
458
  "lo": 0.7000000000000001,
459
459
  "hi": 0.75,
460
460
  "center": 0.7250000000000001,
461
- "calibrated": 0.8984547461368654
461
+ "calibrated": 0.9009009009009009
462
462
  },
463
463
  {
464
464
  "lo": 0.75,
465
465
  "hi": 0.8,
466
466
  "center": 0.775,
467
- "calibrated": 0.9469026548672567
467
+ "calibrated": 0.9009009009009009
468
468
  },
469
469
  {
470
470
  "lo": 0.8,
471
471
  "hi": 0.8500000000000001,
472
472
  "center": 0.8250000000000001,
473
- "calibrated": 0.9577092511013215
473
+ "calibrated": 0.9009009009009009
474
474
  },
475
475
  {
476
476
  "lo": 0.8500000000000001,
477
477
  "hi": 0.9,
478
478
  "center": 0.875,
479
- "calibrated": 0.9694719471947195
479
+ "calibrated": 0.9151515151515152
480
480
  },
481
481
  {
482
482
  "lo": 0.9,
483
483
  "hi": 0.9500000000000001,
484
484
  "center": 0.925,
485
- "calibrated": 0.9838576158940397
485
+ "calibrated": 0.9930601895102096
486
486
  },
487
487
  {
488
488
  "lo": 0.9500000000000001,
package/calibration.json CHANGED
@@ -1,231 +1,243 @@
1
1
  {
2
2
  "model": "neural-weights-en-us",
3
- "model_version": "4.13.0",
3
+ "model_version": "5.3.0",
4
4
  "method": "isotonic-regression (PAVA) over per-span softmax confidence",
5
- "created_from": "data/eval/calibration/confidences-v413.jsonl",
6
- "n_total": 13457,
7
- "n_fit": 10766,
8
- "n_eval": 2691,
5
+ "created_from": "data/eval/calibration/confidences-v530.jsonl",
6
+ "n_total": 35600,
7
+ "n_fit": 28480,
8
+ "n_eval": 7120,
9
9
  "bins": 20,
10
10
  "ece_bins": 15,
11
11
  "metrics": {
12
- "ece_raw_eval": 0.06029901471886912,
13
- "ece_cal_eval": 0.005544698789170681,
14
- "mce_raw_eval": 0.22821355739022853,
15
- "mce_cal_eval": 0.036655484108742864,
16
- "mce_raw_eval_allbins": 0.7814575634642543,
17
- "mce_cal_eval_allbins": 0.6666666666666667,
18
- "brier_raw_eval": 0.028470218500975978,
19
- "brier_cal_eval": 0.023577740980550897,
20
- "ece_raw_oa_eval": 0.05597836419388412,
21
- "ece_cal_oa_eval": 0.019335527504590146,
22
- "ece_raw_corpus_eval": 0.06545452767211968,
23
- "ece_cal_corpus_eval": 0.009526540018422396
12
+ "ece_raw_eval": 0.0677000843149504,
13
+ "ece_cal_eval": 0.002846303809090927,
14
+ "mce_raw_eval": 0.22307581874524562,
15
+ "mce_cal_eval": 0.05785123966942152,
16
+ "mce_raw_eval_allbins": 0.9008025428391669,
17
+ "mce_cal_eval_allbins": 0.3333333333333333,
18
+ "brier_raw_eval": 0.02129048121980521,
19
+ "brier_cal_eval": 0.015551509923503798,
20
+ "ece_raw_oa_eval": 0.06975270093527261,
21
+ "ece_cal_oa_eval": 0.0016543915961919663,
22
+ "ece_raw_corpus_eval": 0.06690828086195735,
23
+ "ece_cal_corpus_eval": 0.0036953442929299886
24
24
  },
25
25
  "per_tag_ece": {
26
26
  "house_number": {
27
- "n": 280,
28
- "acc": 0.875,
29
- "ece_raw": 0.05509278685445104,
30
- "ece_cal": 0.08537893025462967
27
+ "n": 815,
28
+ "acc": 0.9116564417177914,
29
+ "ece_raw": 0.029552022755653974,
30
+ "ece_cal": 0.060257174657242285
31
31
  },
32
32
  "locality": {
33
- "n": 576,
34
- "acc": 0.9583333333333334,
35
- "ece_raw": 0.03092615798253389,
36
- "ece_cal": 0.031805913635991084
33
+ "n": 1570,
34
+ "acc": 0.9802547770700637,
35
+ "ece_raw": 0.047909918029567525,
36
+ "ece_cal": 0.007767010879028461
37
37
  },
38
38
  "postcode": {
39
- "n": 604,
40
- "acc": 0.9867549668874173,
41
- "ece_raw": 0.07691422404152229,
42
- "ece_cal": 0.016352860427164693
39
+ "n": 1570,
40
+ "acc": 1.0,
41
+ "ece_raw": 0.09229924577523453,
42
+ "ece_cal": 0.016096487596793045
43
43
  },
44
44
  "region": {
45
- "n": 451,
46
- "acc": 1.0,
47
- "ece_raw": 0.07172181725109478,
48
- "ece_cal": 0.013513702387324952
45
+ "n": 1244,
46
+ "acc": 0.9991961414790996,
47
+ "ece_raw": 0.07998184691784783,
48
+ "ece_cal": 0.014224405573847645
49
49
  },
50
50
  "street": {
51
- "n": 323,
52
- "acc": 0.9938080495356038,
53
- "ece_raw": 0.09485080545096047,
54
- "ece_cal": 0.027034708166829614
51
+ "n": 828,
52
+ "acc": 0.9927536231884058,
53
+ "ece_raw": 0.07913034025409357,
54
+ "ece_cal": 0.017993784066485708
55
55
  },
56
56
  "street_suffix": {
57
- "n": 247,
57
+ "n": 554,
58
58
  "acc": 1.0,
59
- "ece_raw": 0.09025051456111624,
60
- "ece_cal": 0.02655534057400275
59
+ "ece_raw": 0.07655562851486138,
60
+ "ece_cal": 0.017241106552309277
61
61
  },
62
62
  "venue": {
63
- "n": 141,
64
- "acc": 0.9929078014184397,
65
- "ece_raw": 0.09150430770890858,
66
- "ece_cal": 0.02300676691383495
63
+ "n": 366,
64
+ "acc": 0.9863387978142076,
65
+ "ece_raw": 0.07224547301920636,
66
+ "ece_cal": 0.014302441134295414
67
67
  }
68
68
  },
69
69
  "per_locale_ece": {
70
+ "DE": {
71
+ "n": 185,
72
+ "acc": 0.8918918918918919,
73
+ "ece_raw": 0.08516095367022304,
74
+ "ece_cal": 0.08311326848532197
75
+ },
70
76
  "FR": {
71
- "n": 299,
72
- "acc": 0.9765886287625418,
73
- "ece_raw": 0.07321595575448211,
74
- "ece_cal": 0.008181949744243805
77
+ "n": 804,
78
+ "acc": 0.9626865671641791,
79
+ "ece_raw": 0.06773999694170439,
80
+ "ece_cal": 0.0048981350022959955
81
+ },
82
+ "NL": {
83
+ "n": 205,
84
+ "acc": 1.0,
85
+ "ece_raw": 0.15341190185268108,
86
+ "ece_cal": 0.04619157395115916
75
87
  },
76
88
  "US": {
77
- "n": 2239,
78
- "acc": 0.9785618579723091,
79
- "ece_raw": 0.06228530930926101,
80
- "ece_cal": 0.010679199458733597
89
+ "n": 5926,
90
+ "acc": 0.9866689166385421,
91
+ "ece_raw": 0.06788337090888338,
92
+ "ece_cal": 0.00434667116440722
81
93
  }
82
94
  },
83
95
  "abstention_curve": [
84
96
  {
85
97
  "threshold": 0.5,
86
- "coverage": 0.9996283909327388,
87
- "precision": 0.9739776951672863,
88
- "reviewed": 0.0003716090672611916
98
+ "coverage": 0.9984550561797753,
99
+ "precision": 0.9834013222675482,
100
+ "reviewed": 0.0015449438202247423
89
101
  },
90
102
  {
91
103
  "threshold": 0.8,
92
- "coverage": 0.989223337049424,
93
- "precision": 0.976709241172051,
94
- "reviewed": 0.010776662950576
104
+ "coverage": 0.9955056179775281,
105
+ "precision": 0.9844808126410836,
106
+ "reviewed": 0.004494382022471877
95
107
  },
96
108
  {
97
109
  "threshold": 0.9,
98
- "coverage": 0.967670011148272,
99
- "precision": 0.9796466973886329,
100
- "reviewed": 0.032329988851728
110
+ "coverage": 0.9922752808988764,
111
+ "precision": 0.9851380042462845,
112
+ "reviewed": 0.0077247191011236005
101
113
  },
102
114
  {
103
115
  "threshold": 0.95,
104
- "coverage": 0.8725380899293943,
105
- "precision": 0.9885008517887564,
106
- "reviewed": 0.1274619100706057
116
+ "coverage": 0.9174157303370787,
117
+ "precision": 0.9882118799755052,
118
+ "reviewed": 0.08258426966292132
107
119
  },
108
120
  {
109
121
  "threshold": 0.97,
110
- "coverage": 0.8450390189520625,
111
- "precision": 0.9890061565523307,
112
- "reviewed": 0.15496098104793754
122
+ "coverage": 0.8452247191011236,
123
+ "precision": 0.9910269192422732,
124
+ "reviewed": 0.15477528089887638
113
125
  }
114
126
  ],
115
127
  "reliability_raw": [
116
128
  {
117
129
  "lo": 0.0,
118
130
  "hi": 0.06666666666666667,
119
- "n": 0,
120
- "conf": null,
121
- "acc": null
131
+ "n": 11,
132
+ "conf": 0.00859130897395197,
133
+ "acc": 0.0
122
134
  },
123
135
  {
124
136
  "lo": 0.06666666666666667,
125
137
  "hi": 0.13333333333333333,
126
- "n": 0,
127
- "conf": null,
128
- "acc": null
138
+ "n": 1,
139
+ "conf": 0.09919745716083309,
140
+ "acc": 1.0
129
141
  },
130
142
  {
131
143
  "lo": 0.13333333333333333,
132
144
  "hi": 0.2,
133
- "n": 0,
134
- "conf": null,
135
- "acc": null
145
+ "n": 2,
146
+ "conf": 0.17642642318948423,
147
+ "acc": 0.0
136
148
  },
137
149
  {
138
150
  "lo": 0.2,
139
151
  "hi": 0.26666666666666666,
140
- "n": 1,
141
- "conf": 0.21854243653574568,
152
+ "n": 2,
153
+ "conf": 0.2149186591032387,
142
154
  "acc": 1.0
143
155
  },
144
156
  {
145
157
  "lo": 0.26666666666666666,
146
158
  "hi": 0.3333333333333333,
147
- "n": 1,
148
- "conf": 0.3069032897649658,
149
- "acc": 0.0
159
+ "n": 5,
160
+ "conf": 0.3071612967943654,
161
+ "acc": 0.6
150
162
  },
151
163
  {
152
164
  "lo": 0.3333333333333333,
153
165
  "hi": 0.4,
154
- "n": 2,
155
- "conf": 0.3887903817503637,
156
- "acc": 0.0
166
+ "n": 5,
167
+ "conf": 0.3589545138265823,
168
+ "acc": 0.6
157
169
  },
158
170
  {
159
171
  "lo": 0.4,
160
172
  "hi": 0.4666666666666667,
161
- "n": 4,
162
- "conf": 0.4441922469561045,
163
- "acc": 0.75
173
+ "n": 7,
174
+ "conf": 0.4312700172192481,
175
+ "acc": 0.7142857142857143
164
176
  },
165
177
  {
166
178
  "lo": 0.4666666666666667,
167
179
  "hi": 0.5333333333333333,
168
- "n": 15,
169
- "conf": 0.5044011299509948,
170
- "acc": 0.8666666666666667
180
+ "n": 13,
181
+ "conf": 0.5000625265767356,
182
+ "acc": 0.6923076923076923
171
183
  },
172
184
  {
173
185
  "lo": 0.5333333333333333,
174
186
  "hi": 0.6,
175
- "n": 14,
176
- "conf": 0.5634148099371796,
177
- "acc": 0.6428571428571429
187
+ "n": 12,
188
+ "conf": 0.5644492849667573,
189
+ "acc": 0.9166666666666666
178
190
  },
179
191
  {
180
192
  "lo": 0.6,
181
193
  "hi": 0.6666666666666666,
182
- "n": 22,
183
- "conf": 0.6354228062461351,
184
- "acc": 0.8636363636363636
194
+ "n": 30,
195
+ "conf": 0.6344105236624018,
196
+ "acc": 0.8
185
197
  },
186
198
  {
187
199
  "lo": 0.6666666666666666,
188
200
  "hi": 0.7333333333333333,
189
- "n": 25,
190
- "conf": 0.7012611737412813,
191
- "acc": 0.92
201
+ "n": 48,
202
+ "conf": 0.7016889167234245,
203
+ "acc": 0.9166666666666666
192
204
  },
193
205
  {
194
206
  "lo": 0.7333333333333333,
195
207
  "hi": 0.8,
196
- "n": 54,
197
- "conf": 0.7717951224448251,
198
- "acc": 0.9074074074074074
208
+ "n": 129,
209
+ "conf": 0.7691722432702582,
210
+ "acc": 0.9922480620155039
199
211
  },
200
212
  {
201
213
  "lo": 0.8,
202
214
  "hi": 0.8666666666666667,
203
- "n": 113,
204
- "conf": 0.8392579055758018,
205
- "acc": 0.8761061946902655
215
+ "n": 208,
216
+ "conf": 0.8409492900707859,
217
+ "acc": 0.9423076923076923
206
218
  },
207
219
  {
208
220
  "lo": 0.8666666666666667,
209
221
  "hi": 0.9333333333333333,
210
- "n": 1296,
211
- "conf": 0.9161931717918792,
212
- "acc": 0.9868827160493827
222
+ "n": 4146,
223
+ "conf": 0.9156785010478844,
224
+ "acc": 0.9852870236372407
213
225
  },
214
226
  {
215
227
  "lo": 0.9333333333333333,
216
228
  "hi": 1.0,
217
- "n": 1144,
218
- "conf": 0.9501402430653275,
219
- "acc": 0.9842657342657343
229
+ "n": 2501,
230
+ "conf": 0.9460267604697434,
231
+ "acc": 0.9916033586565374
220
232
  }
221
233
  ],
222
234
  "reliability_cal": [
223
235
  {
224
236
  "lo": 0.0,
225
237
  "hi": 0.06666666666666667,
226
- "n": 0,
227
- "conf": null,
228
- "acc": null
238
+ "n": 11,
239
+ "conf": 0.04040404040404041,
240
+ "acc": 0.0
229
241
  },
230
242
  {
231
243
  "lo": 0.06666666666666667,
@@ -258,9 +270,9 @@
258
270
  {
259
271
  "lo": 0.3333333333333333,
260
272
  "hi": 0.4,
261
- "n": 1,
262
- "conf": 0.3333333333333333,
263
- "acc": 1.0
273
+ "n": 0,
274
+ "conf": null,
275
+ "acc": null
264
276
  },
265
277
  {
266
278
  "lo": 0.4,
@@ -272,9 +284,9 @@
272
284
  {
273
285
  "lo": 0.4666666666666667,
274
286
  "hi": 0.5333333333333333,
275
- "n": 1,
276
- "conf": 0.5,
277
- "acc": 0.0
287
+ "n": 0,
288
+ "conf": null,
289
+ "acc": null
278
290
  },
279
291
  {
280
292
  "lo": 0.5333333333333333,
@@ -286,44 +298,44 @@
286
298
  {
287
299
  "lo": 0.6,
288
300
  "hi": 0.6666666666666666,
289
- "n": 0,
290
- "conf": null,
291
- "acc": null
301
+ "n": 18,
302
+ "conf": 0.6322493224932251,
303
+ "acc": 0.6666666666666666
292
304
  },
293
305
  {
294
306
  "lo": 0.6666666666666666,
295
307
  "hi": 0.7333333333333333,
296
- "n": 6,
297
- "conf": 0.7027027027027027,
298
- "acc": 0.5
308
+ "n": 3,
309
+ "conf": 0.6666666666666666,
310
+ "acc": 0.3333333333333333
299
311
  },
300
312
  {
301
313
  "lo": 0.7333333333333333,
302
314
  "hi": 0.8,
303
- "n": 21,
304
- "conf": 0.7941176470588237,
305
- "acc": 0.8095238095238095
315
+ "n": 0,
316
+ "conf": null,
317
+ "acc": null
306
318
  },
307
319
  {
308
320
  "lo": 0.8,
309
321
  "hi": 0.8666666666666667,
310
- "n": 40,
311
- "conf": 0.8616554841087428,
312
- "acc": 0.825
322
+ "n": 22,
323
+ "conf": 0.8305785123966942,
324
+ "acc": 0.7727272727272727
313
325
  },
314
326
  {
315
327
  "lo": 0.8666666666666667,
316
328
  "hi": 0.9333333333333333,
317
- "n": 259,
318
- "conf": 0.9150005696455284,
319
- "acc": 0.8918918918918919
329
+ "n": 69,
330
+ "conf": 0.9046716558417988,
331
+ "acc": 0.8695652173913043
320
332
  },
321
333
  {
322
334
  "lo": 0.9333333333333333,
323
335
  "hi": 1.0,
324
- "n": 2363,
325
- "conf": 0.9905895070861063,
326
- "acc": 0.988573846804909
336
+ "n": 6997,
337
+ "conf": 0.9842065663919752,
338
+ "acc": 0.9862798342146635
327
339
  }
328
340
  ],
329
341
  "table": [
@@ -331,121 +343,121 @@
331
343
  "lo": 0.0,
332
344
  "hi": 0.05,
333
345
  "center": 0.025,
334
- "calibrated": 0.0
346
+ "calibrated": 0.044444444444444446
335
347
  },
336
348
  {
337
349
  "lo": 0.05,
338
350
  "hi": 0.1,
339
351
  "center": 0.07500000000000001,
340
- "calibrated": 0.0
352
+ "calibrated": 0.6
341
353
  },
342
354
  {
343
355
  "lo": 0.1,
344
356
  "hi": 0.15000000000000002,
345
357
  "center": 0.125,
346
- "calibrated": 0.0
358
+ "calibrated": 0.6
347
359
  },
348
360
  {
349
361
  "lo": 0.15000000000000002,
350
362
  "hi": 0.2,
351
363
  "center": 0.17500000000000002,
352
- "calibrated": 0.3333333333333333
364
+ "calibrated": 0.6341463414634146
353
365
  },
354
366
  {
355
367
  "lo": 0.2,
356
368
  "hi": 0.25,
357
369
  "center": 0.225,
358
- "calibrated": 0.5
370
+ "calibrated": 0.6341463414634146
359
371
  },
360
372
  {
361
373
  "lo": 0.25,
362
374
  "hi": 0.30000000000000004,
363
375
  "center": 0.275,
364
- "calibrated": 0.5
376
+ "calibrated": 0.6341463414634146
365
377
  },
366
378
  {
367
379
  "lo": 0.30000000000000004,
368
380
  "hi": 0.35000000000000003,
369
381
  "center": 0.32500000000000007,
370
- "calibrated": 0.5
382
+ "calibrated": 0.6341463414634146
371
383
  },
372
384
  {
373
385
  "lo": 0.35000000000000003,
374
386
  "hi": 0.4,
375
387
  "center": 0.375,
376
- "calibrated": 0.7027027027027027
388
+ "calibrated": 0.6341463414634146
377
389
  },
378
390
  {
379
391
  "lo": 0.4,
380
392
  "hi": 0.45,
381
393
  "center": 0.42500000000000004,
382
- "calibrated": 0.7027027027027027
394
+ "calibrated": 0.6666666666666666
383
395
  },
384
396
  {
385
397
  "lo": 0.45,
386
398
  "hi": 0.5,
387
399
  "center": 0.475,
388
- "calibrated": 0.7083753763023031
400
+ "calibrated": 0.8181818181818182
389
401
  },
390
402
  {
391
403
  "lo": 0.5,
392
404
  "hi": 0.55,
393
405
  "center": 0.525,
394
- "calibrated": 0.7941176470588235
406
+ "calibrated": 0.8333333333333334
395
407
  },
396
408
  {
397
409
  "lo": 0.55,
398
410
  "hi": 0.6000000000000001,
399
411
  "center": 0.5750000000000001,
400
- "calibrated": 0.8623853211009175
412
+ "calibrated": 0.8333333333333334
401
413
  },
402
414
  {
403
415
  "lo": 0.6000000000000001,
404
416
  "hi": 0.65,
405
417
  "center": 0.625,
406
- "calibrated": 0.8623853211009175
418
+ "calibrated": 0.9047619047619048
407
419
  },
408
420
  {
409
421
  "lo": 0.65,
410
422
  "hi": 0.7000000000000001,
411
423
  "center": 0.675,
412
- "calibrated": 0.8623853211009175
424
+ "calibrated": 0.9047619047619048
413
425
  },
414
426
  {
415
427
  "lo": 0.7000000000000001,
416
428
  "hi": 0.75,
417
429
  "center": 0.7250000000000001,
418
- "calibrated": 0.8769230769230769
430
+ "calibrated": 0.9354346720894764
419
431
  },
420
432
  {
421
433
  "lo": 0.75,
422
434
  "hi": 0.8,
423
435
  "center": 0.775,
424
- "calibrated": 0.917864476386037
436
+ "calibrated": 0.9354346720894764
425
437
  },
426
438
  {
427
439
  "lo": 0.8,
428
440
  "hi": 0.8500000000000001,
429
441
  "center": 0.8250000000000001,
430
- "calibrated": 0.917864476386037
442
+ "calibrated": 0.9354346720894764
431
443
  },
432
444
  {
433
445
  "lo": 0.8500000000000001,
434
446
  "hi": 0.9,
435
447
  "center": 0.875,
436
- "calibrated": 0.917864476386037
448
+ "calibrated": 0.9354346720894764
437
449
  },
438
450
  {
439
451
  "lo": 0.9,
440
452
  "hi": 0.9500000000000001,
441
453
  "center": 0.925,
442
- "calibrated": 0.9924895345973898
454
+ "calibrated": 0.9907799651133815
443
455
  },
444
456
  {
445
457
  "lo": 0.9500000000000001,
446
458
  "hi": 1.0,
447
459
  "center": 0.9750000000000001,
448
- "calibrated": 0.9924895345973898
460
+ "calibrated": 0.9907799651133815
449
461
  }
450
462
  ]
451
463
  }
package/model-card.json CHANGED
@@ -1,8 +1,8 @@
1
1
  {
2
2
  "name": "neural-weights-en-us",
3
- "version": "5.3.0",
4
- "model_lineage": "v2.2.0-full-family / step 100000 — the FIRST ground-up from-scratch train on the spliced vocab (v0.8-era schedule: lr 1.5e-4, warmup 1000, batch 128; corpus v0.10.0 = 698 shards incl. the four-orthography leading-name-boundary family @6.0: fr-bare-street #831 / si-bare-village / no-street-led / cz-pcfirst-preposition; tokenizer v0.7.1-nsplice UNCHANGED — model-only bump). Archived NO-PROMOTE 2026-07-03 under the pre-#942 calculus (SI sentinels model-level fail); PROMOTED 2026-07-04 after the same-stack re-gate under the #942 floor (operator fork on #901). THE HEADLINE: fixes #949 — the shipped v5.2.0 pair fragmented FR streets bare AND anchored (16/40, 13/40 on the standing 40-row harness; 'Allée'→'e' at the é piece boundary; ungated because no fr.street floor existed) — v2.2.0 reads 34/40 bare / 37/40 anchored. Re-gate: US-2k ni PASS (p50 3.27 IDENTICAL), FR-3k ni PASS deconfounded (1.30km both), CZ wrong-city 13.7→6.6%, SK 13.3→6.6%, PL 7.1→6.2%; SE +4.6pp resolve (90.2→94.8), Åbo 0.12km vs 859km. TRADES (operator-approved, stated): SI wrong-city 8.7→11.4% (+2.7pp) / resolve −1.5pp — the composition-insensitive no-street class, floored model-independently by postalCompoundRecovery (#942 default-ON: candidate+floor recovered ALL 55 composition-lost rows, SI 92.2→100% resolve on the insurance leg); fr.postcode 99.2 (floor revised 99.3→99.2, scripts/eval/gates/v5.3.0-family.json, stated reason); int8 homograph quantization delta 2.4 vs the 1.5 cap (fp32 passes every floor; int8 homograph 85.1 clears its 83.3 floor) — per-artifact exception, cap unchanged. PRIOR LINEAGE (the v5.2.0 base this replaces): v5.1.0 bsplice-expanded (UNCHANGED encoder — see below) + the #912-lever-4 NORDIC vocab-splice with FVT mean-init embeddings (the same training-free #884 mechanism, second application): +5,331 Nordic diacritic pieces from OA fi/se/no/dk/is STREET/CITY text + the gazetteer ALIAS names (the exonym feed — 'Åbo' is Swedish for Turku and appears in no native OA text; without alias mass the piece is never minted and the decode drop survives, measured as the 134km v1 residual). Tokenizer v0.7.1-nsplice = v0.6.0-bsplice + the Nordic pieces; English tokenization byte-identical; US resolution byte-identical 2000/2000. Gate: 14/14 per-locale ni PASS (incl. the five #900-accepted acute-vowel overlap locales CZ/ES/PL/SI/SK), SE resolve 83.8→90.2% (p90res 113→12.7km), NO 92.7→98.1%, DK 91.4→95.2%, Åbo→Turku 0.12km. PRIOR LINEAGE: v1.9.3a3-anchor-absorption encoder (UNCHANGED, step 80000 — see the v4.15.0 lineage below) + the #884 UNIGRAM vocab-splice with FVT mean-init embeddings — the training-free CZ/PL/SK/SI diacritic fix. The 48k SentencePiece vocab carried the diacritic CHARACTERS but no multi-char subwords containing them, so every diacritic isolated its own piece (Vysoká → [▁V, ys, ok, á], ~3.3× English fertility) and the fragmented span resolved to the WRONG CITY. A shard retrain (v196) was falsified (held US, regressed CZ wrong-city 44→58% at convergence). The fix: train a CZ/PL SP unigram, splice ONLY its diacritic-bearing pieces into the vocab (48,000 → 58,582; a diacritic piece can never match an ASCII span, so English tokenizes byte-identically BY CONSTRUCTION — asserted 0/2000), and FVT mean-init the 10,582 new embedding rows from their old-tokenizer constituents. The encoder is byte-for-byte v1.9.3a3; a 2k fine-tune was ablated and retired as slightly harmful. Slovak + Slovenian ride the same pieces for free (shared codepoints + morphology). PRIOR LINEAGE: v1.9.3a3-anchor-absorption / step 80000 — the #220/#723 postcode-anchor-pollution fix on top of v1.9.2-multilocale-au (shipped v4.14.0). Pre-seeded from the v1.9.2 step-040000 checkpoint, continued to 80k on A100. TWO coordinated levers: (1) anchor_paint_mode=shaped — the postcode anchor is painted at TRAIN time on the same postcode-SHAPED spans inference fires on (neural/postcode-anchor.ts collectMatches), closing the #723 WHERE gap (training only ever saw the anchor on GOLD postcode spans, so the model never learned to OVERRIDE a present anchor when a leading 5-digit token is a house number that looks like a ZIP); (2) the synth-anchor-absorption counter-shard (source-weight 6.0) carrying the locality discriminator — leading-5-digit + trailing-postcode → house_number, + locality → house_number, + NO locality + state → postcode (the VT E911 format). RESULT: SLICE-H anchor-ON 20→100 (the #723 leading-5-digit house# bug fixed) at ZERO coordinate cost, confirmed by the non-circular oa-resolver-eval (2000 real OA-US points, full ship-config, v192 vs v1.9.3a3): whole-US locality-match 97.7→97.8 (+0.1), region 99.9 flat, coord p50 3.4 flat, p90 10.7 flat, p99 275.3→259.1; VT — the E911 leading-postcode state where the entire trade lives — FLAT at the coordinate (loc 92.6=92.6), SD +1.0. Trades us.postcode LABEL-F1 −2 (95.0) + fr.postcode −0.2 (99.3), both coordinate-invisible; gate revised (scripts/eval/gates/v4.15.0-boundary.json) with a stated reason, operator-approved. mask-regression gate (#718) PASS; int8↔fp32 delta 0. Ship-config (requires) UNCHANGED from v1.9.2.",
5
- "phase": "v5.3.0 v2.2.0-full-familyfrom-scratch on nsplice; fixes the #949 FR street fragmentation (bare 16→34/40), halves CZ/SK wrong-city (13.7→6.6 / 13.3→6.6), SE +4.6pp, Åbo fixed; SI no-street trade floored by #942",
3
+ "version": "5.4.0",
4
+ "model_lineage": "v2.3.0-nl-postcode — init_from FINE-TUNE of v5.3.0 (v2.2.0-full-family) step-100000 + synth-nl-postcode@6.0, the four leading-name-boundary family shards pinned @12.0 (fr-bare-street/si-village/no-street-led/cz-pcfirst). LR 1e-5, warmup 1000, 5000 steps (~0.7 A100-h fine-tune vs a 3.6h from-scratch — the DeepSeek-signed recipe: the 'any fine-tune tips the #901 family' lore is surgery-lineage-specific, and v5.3.0 is from-scratch so it took the pinned fine-tune WITHOUT tipping the sentinels). THE FIX (#924): the model read the digits-first Dutch postcode '1012 LG' as house_number(1012)+street(LG), pulling the locality into the US situs tier ('1012 LG Amsterdam' -> Amsterdam, NY). Now parses postcode '1012 LG' natively (NL failing-shapes 27->29/31; postcode-tag F1 1.000). Gains beyond NL: US inverse-risk (4-digit house# NOT misread as postcode) 9->10/10 — the fine-tune even fixed a pre-existing leak; FR bare-street chevaleret PARSE fixed (locality=Paris, street='Rue du Chevaleret' vs v5.3.0's fragmentation) — the chevaleret metamorphic gauntlet rows reclassified #829 (resolver case-norm exposed by the better parse, not a regression). GATE: 17/17 per-tag floors PASS (us.postcode 96.2, homograph 85.1, micro 86.8, street 82.1, fr.postcode 99.3), fr.bare_street_intact 85% (#949 leg), mask-regression PASS, arena 79, gauntlet PASS (5 #901 sentinels HELD). macro_f1 0.7613->0.7558 (noise). CARRIED int8-homograph exception: int8<->fp32 delta 2.4 vs the 1.5 cap (fp32 passes every floor; int8 homograph 85.1 clears its 83.3 floor) — the same per-artifact exception v5.3.0 shipped with; cap unchanged for the next candidate. CALIBRATION: v5.3.0's isotonic tables carried (opt-in; the fine-tune's confidence shift is small; re-fit is a follow-up). PRIOR LINEAGE: v2.2.0-full-family / step 100000 — the FIRST ground-up from-scratch train on the spliced vocab (v0.8-era schedule: lr 1.5e-4, warmup 1000, batch 128; corpus v0.10.0 = 698 shards incl. the four-orthography leading-name-boundary family @6.0: fr-bare-street #831 / si-bare-village / no-street-led / cz-pcfirst-preposition; tokenizer v0.7.1-nsplice UNCHANGED — model-only bump). Archived NO-PROMOTE 2026-07-03 under the pre-#942 calculus (SI sentinels model-level fail); PROMOTED 2026-07-04 after the same-stack re-gate under the #942 floor (operator fork on #901). THE HEADLINE: fixes #949 — the shipped v5.2.0 pair fragmented FR streets bare AND anchored (16/40, 13/40 on the standing 40-row harness; 'Allée'→'e' at the é piece boundary; ungated because no fr.street floor existed) — v2.2.0 reads 34/40 bare / 37/40 anchored. Re-gate: US-2k ni PASS (p50 3.27 IDENTICAL), FR-3k ni PASS deconfounded (1.30km both), CZ wrong-city 13.7→6.6%, SK 13.3→6.6%, PL 7.1→6.2%; SE +4.6pp resolve (90.2→94.8), Åbo 0.12km vs 859km. TRADES (operator-approved, stated): SI wrong-city 8.7→11.4% (+2.7pp) / resolve −1.5pp — the composition-insensitive no-street class, floored model-independently by postalCompoundRecovery (#942 default-ON: candidate+floor recovered ALL 55 composition-lost rows, SI 92.2→100% resolve on the insurance leg); fr.postcode 99.2 (floor revised 99.3→99.2, scripts/eval/gates/v5.3.0-family.json, stated reason); int8 homograph quantization delta 2.4 vs the 1.5 cap (fp32 passes every floor; int8 homograph 85.1 clears its 83.3 floor) — per-artifact exception, cap unchanged. PRIOR LINEAGE (the v5.2.0 base this replaces): v5.1.0 bsplice-expanded (UNCHANGED encoder — see below) + the #912-lever-4 NORDIC vocab-splice with FVT mean-init embeddings (the same training-free #884 mechanism, second application): +5,331 Nordic diacritic pieces from OA fi/se/no/dk/is STREET/CITY text + the gazetteer ALIAS names (the exonym feed — 'Åbo' is Swedish for Turku and appears in no native OA text; without alias mass the piece is never minted and the decode drop survives, measured as the 134km v1 residual). Tokenizer v0.7.1-nsplice = v0.6.0-bsplice + the Nordic pieces; English tokenization byte-identical; US resolution byte-identical 2000/2000. Gate: 14/14 per-locale ni PASS (incl. the five #900-accepted acute-vowel overlap locales CZ/ES/PL/SI/SK), SE resolve 83.8→90.2% (p90res 113→12.7km), NO 92.7→98.1%, DK 91.4→95.2%, Åbo→Turku 0.12km. PRIOR LINEAGE: v1.9.3a3-anchor-absorption encoder (UNCHANGED, step 80000 — see the v4.15.0 lineage below) + the #884 UNIGRAM vocab-splice with FVT mean-init embeddings — the training-free CZ/PL/SK/SI diacritic fix. The 48k SentencePiece vocab carried the diacritic CHARACTERS but no multi-char subwords containing them, so every diacritic isolated its own piece (Vysoká → [▁V, ys, ok, á], ~3.3× English fertility) and the fragmented span resolved to the WRONG CITY. A shard retrain (v196) was falsified (held US, regressed CZ wrong-city 44→58% at convergence). The fix: train a CZ/PL SP unigram, splice ONLY its diacritic-bearing pieces into the vocab (48,000 → 58,582; a diacritic piece can never match an ASCII span, so English tokenizes byte-identically BY CONSTRUCTION — asserted 0/2000), and FVT mean-init the 10,582 new embedding rows from their old-tokenizer constituents. The encoder is byte-for-byte v1.9.3a3; a 2k fine-tune was ablated and retired as slightly harmful. Slovak + Slovenian ride the same pieces for free (shared codepoints + morphology). PRIOR LINEAGE: v1.9.3a3-anchor-absorption / step 80000 — the #220/#723 postcode-anchor-pollution fix on top of v1.9.2-multilocale-au (shipped v4.14.0). Pre-seeded from the v1.9.2 step-040000 checkpoint, continued to 80k on A100. TWO coordinated levers: (1) anchor_paint_mode=shaped — the postcode anchor is painted at TRAIN time on the same postcode-SHAPED spans inference fires on (neural/postcode-anchor.ts collectMatches), closing the #723 WHERE gap (training only ever saw the anchor on GOLD postcode spans, so the model never learned to OVERRIDE a present anchor when a leading 5-digit token is a house number that looks like a ZIP); (2) the synth-anchor-absorption counter-shard (source-weight 6.0) carrying the locality discriminator — leading-5-digit + trailing-postcode → house_number, + locality → house_number, + NO locality + state → postcode (the VT E911 format). RESULT: SLICE-H anchor-ON 20→100 (the #723 leading-5-digit house# bug fixed) at ZERO coordinate cost, confirmed by the non-circular oa-resolver-eval (2000 real OA-US points, full ship-config, v192 vs v1.9.3a3): whole-US locality-match 97.7→97.8 (+0.1), region 99.9 flat, coord p50 3.4 flat, p90 10.7 flat, p99 275.3→259.1; VT — the E911 leading-postcode state where the entire trade lives — FLAT at the coordinate (loc 92.6=92.6), SD +1.0. Trades us.postcode LABEL-F1 −2 (95.0) + fr.postcode −0.2 (99.3), both coordinate-invisible; gate revised (scripts/eval/gates/v4.15.0-boundary.json) with a stated reason, operator-approved. mask-regression gate (#718) PASS; int8↔fp32 delta 0. Ship-config (requires) UNCHANGED from v1.9.2.",
5
+ "phase": "v5.4.0 / v2.3.0-nl-postcode — the #924 NL digits-first postcode fix as a pinned init_from fine-tune; sentinels held, US inverse-risk + FR chevaleret parse improved",
6
6
  "license": "AGPL-3.0-only OR LicenseRef-Commercial",
7
7
  "locale": "en-us",
8
8
  "training": {
@@ -122,9 +122,9 @@
122
122
  },
123
123
  "calibration": {
124
124
  "method": "isotonic-regression (PAVA) over per-span softmax confidence; OPT-IN via core/decoder createCalibrator",
125
- "held_out_ece_raw": 0.0673,
126
- "held_out_ece_calibrated": 0.0035,
127
- "note": "calibration.json + calibration-per-locale.json carried forward from v4.4.0 (architecture + label space unchanged across v1.5.0→v1.8.0). Re-fit recommended next train. Apply via @mailwoman/core/decoder's createCalibrator; default parse output is byte-stable when omitted."
125
+ "held_out_ece_raw": 0.0698,
126
+ "held_out_ece_calibrated": 0.0017,
127
+ "note": "RE-FIT on v5.3.0 (2026-07-04): calibration.json + calibration-per-locale.json are now fitted on THIS model's confidence distribution (35,600 spans over the 8k held-out set), not carried from v4.x. The from-scratch model was under-confident (meanConf 0.913 vs acc 0.980, gap -0.067); the fit corrects held-out OA ECE 0.0698→0.0017 (overall 0.06770.0028, per-locale NL 0.166→0.000 / DE 0.114→0.003). Reproduce: collect-span-confidences.ts → fit-isotonic-calibration.py --model-version 5.3.0. Apply via @mailwoman/core/decoder's createCalibrator; default parse output is byte-stable when omitted."
128
128
  },
129
129
  "base_relpath": "/mnt/playpen/mailwoman-data/models/candidates/v220-full-family (fp32 + int8; training output on Modal volume)",
130
130
  "eval": {
package/model.onnx CHANGED
Binary file
package/package.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "@mailwoman/neural-weights-en-us",
3
- "version": "5.3.0",
3
+ "version": "5.4.0",
4
4
  "description": "Mailwoman neural-classifier weights for locale 'en-us'. Data-only package — loaded by @mailwoman/neural at runtime.",
5
5
  "license": "AGPL-3.0-only OR LicenseRef-Commercial",
6
6
  "repository": {