@mailwoman/neural-weights-en-us 4.4.0 → 4.5.0

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.
Files changed (3) hide show
  1. package/model-card.json +37 -39
  2. package/model.onnx +0 -0
  3. package/package.json +1 -1
package/model-card.json CHANGED
@@ -1,17 +1,17 @@
1
1
  {
2
2
  "name": "neural-weights-en-us",
3
- "version": "4.4.0",
4
- "model_lineage": "v1.3.0-boundary-consolidation / step 40000 — from scratch on corpus v0.4.15-boundary: the v4.3.0 recipe + three probe-passed levers (glue augmentation #513, real-TIGER intersection shard #487, codex po_box/cedex coverage shard), old synth-po-box retired. Ships with TWO declared inference behaviors: addressSystemConventions:'auto' (v4.3.0) + bridgePunctuationGaps:true (the span bridge — the corpus label format cannot express punctuation inside spans; the bridge merges same-tag fragments across intra-token punctuation). Gate: docs/articles/evals/2026-06-11-v4.4.0-ship-gate.md; tokenizer 0.6.0-a0",
5
- "phase": "Stage 3 — v1.3 boundary: the last starved tags alive (po_box/cedex/intersections) + the arena debt closed",
3
+ "version": "4.5.0",
4
+ "model_lineage": "v1.4.0-charoffset / step 40000 — the FIRST char-offset-corpus model (#519). From scratch on corpus v0.5.0 (676.6M rows: a from-source rebuild in the char-offset span-label format + the 7 re-emitted parity overlays). IDENTICAL recipe to v4.4.0's v1.3.0-boundary-consolidation — the ONLY variable is the label format (token BIO → char-offset span triple), so the v4.4.0 floors are the hold-the-line contract. HEADLINE: the span bridge (bridgePunctuationGaps) is RETIRED — the char-offset format lets the model learn dotted po_box spans intrinsically (po_box 90.0 graded bridge-OFF, ≥ the 89.1 the bridge achieved), and bridge-on vs bridge-off is byte-identical on every tag (confirmed no-op). One declared inference behavior remains: addressSystemConventions:'auto'. Gate: scripts/eval/gates/v0.5.0-bridge.json + docs/articles/evals/2026-06-12-v050-charoffset-launch.md; tokenizer 0.6.0-a0.",
5
+ "phase": "v1.4 char-offset — the structural cure: span-label format ships, the decode-side span bridge retired (po_box learned intrinsically)",
6
6
  "license": "AGPL-3.0-only",
7
7
  "locale": "en-us",
8
8
  "training": {
9
- "corpus_version": "0.4.15-boundary (v0.4.12 + intersection + po_box/cedex shards; loader relabel + glue augmentation)",
9
+ "corpus_version": "0.5.0 (char-offset span labels #519: from-source base + 7 re-emitted parity overlays)",
10
10
  "tokenizer_version": "0.6.0-a0",
11
11
  "steps": 40000,
12
12
  "best_step": 40000,
13
13
  "hardware": "NVIDIA A100-SXM4-40GB (Modal cloud)",
14
- "recipe": "v1.3.0-boundary-consolidation: from-scratch 40k, augment_glue_prob 0.25, synth-intersection 1.0, synth-po-box-cedex 1.5 (old synth-po-box retired, stated), affix relabel pass, gazetteer anchor + choreography, CE-only, lr=1.5e-4, seed 42."
14
+ "recipe": "v1.4.0-charoffset: from-scratch 40k on the char-offset corpus, recipe held identical to v1.3.0-boundary (augment_glue_prob 0.25, synth-intersection 1.0, synth-po-box-cedex 1.5, affix relabel, gazetteer anchor + choreography, CE-only, lr=1.5e-4, seed 42) — the format is the single variable."
15
15
  },
16
16
  "architecture": {
17
17
  "hidden_size": 384,
@@ -21,7 +21,7 @@
21
21
  "max_position_embeddings": 128,
22
22
  "vocab_size": 48000,
23
23
  "num_labels": 33,
24
- "params": "29.3M (29M encoder + 9M embedding from 48K vocab)",
24
+ "params": "29.6M (29M encoder + 9M embedding from 48K vocab)",
25
25
  "crf_at_training": false,
26
26
  "crf_at_inference": true,
27
27
  "phrase_priors": true
@@ -79,9 +79,9 @@
79
79
  "intersection_a",
80
80
  "intersection_b"
81
81
  ],
82
- "notes": "v4.4.0 — the boundary consolidation: po_box 0→89.1, cedex 0→96.1, intersections 0→100 (real-OOD, P≥95) — the parity scorecard's last empty rows filled; the v4.3.0 perturb-arena debt closed (64→72, gated). The ship config REQUIRES bridgePunctuationGaps:true + addressSystemConventions:'auto' — without the bridge, dotted surfaces (P.O. Box, C.P.) decode as period-truncated fragments (the corpus alignment tokenizer cannot label punctuation; measured: po_box 60.4 without, 89.1 with). Bridge merges intra-token punctuation gaps ONLY (commas excluded — they separate genuine spans). Known follow-ups: char-offset corpus labels (the structural cure), conventions loss-mask rides the next run, FR region tail 25.6 (floored, recovering).",
82
+ "notes": "v4.5.0 — the char-offset cure. The corpus now labels spans by character offset (#519), so the model learns dotted surfaces (P.O. Box, C.P.) as contiguous spans WITHOUT the decode-side span bridge: po_box 90.0 graded bridge-OFF (v4.4.0 needed the bridge to reach 89.1). The bridge is now a CONFIRMED NO-OP (bridge-on == bridge-off on every tag) and is retired — it may be left enabled harmlessly. 15/17 tags hold flat-or-better actual-vs-v4.4.0. KNOWN REGRESSION (accepted trade, recover-next-run): fr.house_number 97.7 → ~89.6 (−8pp). Root cause #560 — the model is order-blind on FR house_numbers in REVERSED (postcode-first) order (\"47110 Locality, 69 Street\"), which the FR training (BAN, canonical-order) never taught; v4.4.0's 97.7 was bridge-rescued on exactly these. The FR house_number eval is 85% one reversed-order cluster, so the metric is hypersensitive. Fix queued: a reversed-order FR synth shard (the German both-order shape). Other follow-up: corpus-v0.5.1 code-point offset re-align (the #558 astral-skip's lasting fix).",
83
83
  "format": {
84
- "model": "ONNX int8 dynamic (quantized from fp32)",
84
+ "model": "ONNX int8 dynamic (quantized from fp32, max fp32↔int8 delta 0.3pp)",
85
85
  "tokenizer": "SentencePiece unigram, byte_fallback=true, vocab_size=48000",
86
86
  "max_sequence_length": 128,
87
87
  "opset": 17,
@@ -99,53 +99,51 @@
99
99
  "method": "isotonic-regression (PAVA) over per-span softmax confidence; OPT-IN via core/decoder createCalibrator",
100
100
  "held_out_ece_raw": 0.0673,
101
101
  "held_out_ece_calibrated": 0.0035,
102
- "note": "calibration.json is the global table; calibration-per-locale.json carries per-locale tables (the global table under-serves DE/NL). Apply via @mailwoman/core/decoder's createCalibrator; default parse output is byte-stable when omitted."
102
+ "note": "calibration.json + calibration-per-locale.json carried forward from v4.4.0 (architecture + label space unchanged). Re-fit recommended next train. Apply via @mailwoman/core/decoder's createCalibrator; default parse output is byte-stable when omitted."
103
103
  },
104
- "base_relpath": "/data/output-v097-unit-v3-s42/checkpoints/step-020000",
104
+ "base_relpath": "/data/output-v140-charoffset-s42/checkpoints/step-040000",
105
105
  "eval": {
106
- "ship_gate_2026_06_11_night": {
107
- "promotion_gate": "PASS 17/17 (gates/v4.4.0-boundary.json, int8-graded, max delta 0.3pp; three-battery record in the gate doc)",
108
- "honest_eval_vt": {
109
- "n": 1428,
110
- "region_match_pct": 99.6,
111
- "coord_p50_km": 3.4,
112
- "coord_p90_km": 7.5,
113
- "verdict": "PASS"
114
- },
115
- "arena_perturb_pct": 72.0
106
+ "gate_2026_06_12_charoffset": {
107
+ "promotion_gate": "FAIL 16/17 (gates/v0.5.0-bridge.json, bridge-OFF) — the single miss is fr.house_number 89.6 vs floor 91 (accepted trade, #560). int8-vs-fp32 max delta 0.3pp (≤1.5 cap). bridge-retirement VALIDATED: po_box 90.0 bridge-OFF ≥ 89.1.",
108
+ "bridge_retired": true,
109
+ "gate_doc": "docs/articles/evals/2026-06-12-v050-charoffset-launch.md"
116
110
  },
117
111
  "per_component_int8_shipconfig": {
118
112
  "us": {
119
- "postcode": 98.3,
120
- "country_homograph": 89.8,
121
- "micro": 86.1,
122
- "locality": 75.7,
123
- "region": 90.3,
124
- "street": 77.9,
125
- "street_prefix": 93.6,
126
- "street_suffix": 96.6,
127
- "unit": 92.1,
128
- "po_box_real": 89.1,
113
+ "postcode": 98.5,
114
+ "country_homograph": 87.5,
115
+ "micro": 85.4,
116
+ "locality": 74.0,
117
+ "region": 89.5,
118
+ "street": 75.8,
119
+ "street_prefix": 98.0,
120
+ "street_suffix": 94.9,
121
+ "unit": 97.0,
122
+ "po_box_real": 90.0,
129
123
  "intersection_real": 100.0
130
124
  },
131
125
  "fr": {
132
- "postcode": 99.6,
133
- "house_number": 97.2,
134
- "region": 25.6,
135
- "cedex_real": 96.1
126
+ "postcode": 99.7,
127
+ "house_number": 89.3,
128
+ "region": 35.2,
129
+ "cedex_real": 96.6
136
130
  },
137
131
  "de": {
138
132
  "native_locality_anchor_on": 91.0
139
133
  }
140
134
  },
141
- "known_regressions_vs_4_3_0": {
142
- "fr_postcode": -0.1,
143
- "fr_house_number": -0.5,
144
- "note": "both within single-row noise on their evals; every floor met"
135
+ "vs_4_4_0_actual": {
136
+ "po_box_real": "+0.9 (89.1 bridged → 90.0 intrinsic, bridge RETIRED)",
137
+ "unit_real": "+4.9",
138
+ "country_homograph": "+2.4",
139
+ "street_prefix": "+4.4",
140
+ "fr_region": "+9.6",
141
+ "fr_house_number": "-8.1 (97.7 → 89.6) — KNOWN, #560, reversed-order FR (recover next run)",
142
+ "note": "15/17 tags flat-or-better actual-vs-actual; fr.house_number is the one real casualty (accepted trade)."
145
143
  }
146
144
  },
147
145
  "files_md5": {
148
- "model.onnx": "f086951a807b35e1ef700c0c2662a088",
146
+ "model.onnx": "25f3956de77a252bb9440907eb5b2a37",
149
147
  "tokenizer.model": "b6137e8c52914c9715374268ecaa4bc6"
150
148
  }
151
149
  }
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": "4.4.0",
3
+ "version": "4.5.0",
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",
6
6
  "repository": {