@mailwoman/neural-weights-fr-fr 2.0.0 → 2.2.0

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 (5) hide show
  1. package/README.md +16 -13
  2. package/model-card.json +102 -50
  3. package/model.onnx +0 -0
  4. package/package.json +18 -18
  5. package/tokenizer.model +0 -0
package/README.md CHANGED
@@ -1,30 +1,33 @@
1
1
  # @mailwoman/neural-weights-fr-fr
2
2
 
3
- Phase 2 / Stage 1 (coarse) Mailwoman neural-classifier weights.
3
+ Stage 2 (coarse + venue/street/house_number) Mailwoman neural-classifier weights.
4
4
 
5
5
  - locale: **fr-fr**
6
- - corpus: **0.2.0**
7
- - training steps: **50000**
6
+ - corpus: **0.3.0**
7
+ - training steps: **2200**
8
8
  - hardware: **AMD Radeon 780M (gfx1103) bf16 ~14.6 GiB GTT**
9
9
 
10
- ## Phase 2 §6 status
10
+ ## Per-component F1 targets
11
11
 
12
- **⚠ Below Phase 2 §6 targets (≥95% F1):**
12
+ **⚠ Below per-component F1 targets:**
13
13
 
14
- - `country` F1 = **0.0000** (target ≥0.95)
15
- - `region` F1 = **0.8293** (target ≥0.95)
16
- - `locality` F1 = **0.6471** (target ≥0.95)
17
- - `postcode` F1 = **0.8594** (target ≥0.95)
14
+ - `country` F1 = **0.2112** (target ≥0.95)
15
+ - `region` F1 = **0.1883** (target ≥0.95)
16
+ - `locality` F1 = **0.2736** (target ≥0.95)
17
+ - `postcode` F1 = **0.6916** (target ≥0.95)
18
+ - `venue` F1 = **0.3886** (target ≥0.60)
19
+ - `street` F1 = **0.3016** (target ≥0.70)
20
+ - `house_number` F1 = **0.7866** (target ≥0.80)
18
21
 
19
22
  ## Eval (golden set)
20
23
 
21
- - entries: **74**
22
- - full-parse exact match: **0.5270**
23
- - mean token confidence: **0.9745**
24
+ - entries: **4535**
25
+ - full-parse exact match: **0.0818**
26
+ - mean token confidence: **0.8063**
24
27
 
25
28
  ## Components supported
26
29
 
27
- Stage 1 ships coarse-only: country / region / locality / dependent_locality / postcode / subregion / cedex. Street- and venue-level components are explicit future phases.
30
+ Stage 2 ships coarse (country / region / locality / dependent_locality / postcode / subregion / cedex) plus fine-grained venue / street / house_number. Token classifier emits 21 BIO labels.
28
31
 
29
32
  ## Files
30
33
 
package/model-card.json CHANGED
@@ -1,53 +1,87 @@
1
1
  {
2
2
  "name": "neural-weights-fr-fr",
3
- "version": "0.2.0",
4
- "phase": "Stage 1 (coarse)",
3
+ "version": "0.4.0",
4
+ "phase": "Stage 2 (coarse + venue/street/house_number)",
5
5
  "license": "AGPL-3.0-only",
6
6
  "locale": "fr-fr",
7
7
  "training": {
8
- "corpus_version": "0.2.0",
8
+ "corpus_version": "0.3.0",
9
9
  "tokenizer_version": "0.1.0",
10
- "steps": 50000,
10
+ "steps": 2200,
11
11
  "hardware": "AMD Radeon 780M (gfx1103) bf16 ~14.6 GiB GTT",
12
- "duration_seconds": 23520.0,
12
+ "duration_seconds": 1146.0,
13
13
  "started_at": null,
14
- "completed_at": "2026-05-18T21:33:27.380730Z"
14
+ "completed_at": "2026-05-23T06:21:51.190078Z"
15
15
  },
16
- "components_supported": ["country", "region", "locality", "dependent_locality", "postcode", "subregion", "cedex"],
16
+ "components_supported": [
17
+ "country",
18
+ "region",
19
+ "locality",
20
+ "dependent_locality",
21
+ "postcode",
22
+ "subregion",
23
+ "cedex",
24
+ "venue",
25
+ "street",
26
+ "house_number"
27
+ ],
28
+ "labels": [
29
+ "O",
30
+ "B-country",
31
+ "I-country",
32
+ "B-region",
33
+ "I-region",
34
+ "B-locality",
35
+ "I-locality",
36
+ "B-dependent_locality",
37
+ "I-dependent_locality",
38
+ "B-postcode",
39
+ "I-postcode",
40
+ "B-subregion",
41
+ "I-subregion",
42
+ "B-cedex",
43
+ "I-cedex",
44
+ "B-venue",
45
+ "I-venue",
46
+ "B-street",
47
+ "I-street",
48
+ "B-house_number",
49
+ "I-house_number"
50
+ ],
17
51
  "eval": {
18
- "n_entries": 74,
19
- "full_parse_exact_match": 0.527027027027027,
20
- "mean_token_confidence": 0.974534777700901,
52
+ "n_entries": 4535,
53
+ "full_parse_exact_match": 0.08180815876515987,
54
+ "mean_token_confidence": 0.8062812768727202,
21
55
  "per_component": {
22
56
  "country": {
23
- "precision": 0.0,
24
- "recall": 0.0,
25
- "f1": 0.0,
26
- "support": 6
57
+ "precision": 0.21428571428481394,
58
+ "recall": 0.20816326530527282,
59
+ "f1": 0.21118012372283307,
60
+ "support": 245
27
61
  },
28
62
  "region": {
29
- "precision": 0.8499999999858334,
30
- "recall": 0.80952380951096,
31
- "f1": 0.8292682921697403,
32
- "support": 63
63
+ "precision": 0.342951360263526,
64
+ "recall": 0.129797191887635,
65
+ "f1": 0.18832050661831204,
66
+ "support": 3205
33
67
  },
34
68
  "locality": {
35
- "precision": 0.6874999999892578,
36
- "recall": 0.6111111111026234,
37
- "f1": 0.6470588230216262,
38
- "support": 72
69
+ "precision": 0.24782398452605223,
70
+ "recall": 0.30533214179317686,
71
+ "f1": 0.2735886822759171,
72
+ "support": 3357
39
73
  },
40
74
  "dependent_locality": {
41
- "precision": 0.0,
42
- "recall": 0.0,
43
- "f1": 0.0,
44
- "support": 1
75
+ "precision": 0.005044136191670815,
76
+ "recall": 0.0999999999975,
77
+ "f1": 0.009603841445164863,
78
+ "support": 40
45
79
  },
46
80
  "postcode": {
47
- "precision": 0.8730158730020157,
48
- "recall": 0.8461538461408283,
49
- "f1": 0.8593749994866943,
50
- "support": 65
81
+ "precision": 0.8323890462696731,
82
+ "recall": 0.591610738254835,
83
+ "f1": 0.6916437813892687,
84
+ "support": 2980
51
85
  },
52
86
  "subregion": {
53
87
  "precision": 0.0,
@@ -60,6 +94,24 @@
60
94
  "recall": 0.0,
61
95
  "f1": 0.0,
62
96
  "support": 1
97
+ },
98
+ "venue": {
99
+ "precision": 0.37649063032335905,
100
+ "recall": 0.4014532243411431,
101
+ "f1": 0.38857142807160183,
102
+ "support": 1101
103
+ },
104
+ "street": {
105
+ "precision": 0.35594795539016916,
106
+ "recall": 0.26161202185783416,
107
+ "f1": 0.3015748026611547,
108
+ "support": 2928
109
+ },
110
+ "house_number": {
111
+ "precision": 0.7446153846150028,
112
+ "recall": 0.8335246842704744,
113
+ "f1": 0.7865655466300883,
114
+ "support": 1742
63
115
  }
64
116
  },
65
117
  "calibration": [
@@ -72,56 +124,56 @@
72
124
  {
73
125
  "low": 0.1,
74
126
  "high": 0.2,
75
- "n": 0,
76
- "acc": 0.0
127
+ "n": 21,
128
+ "acc": 0.23809523809523808
77
129
  },
78
130
  {
79
131
  "low": 0.2,
80
132
  "high": 0.3,
81
- "n": 0,
82
- "acc": 0.0
133
+ "n": 666,
134
+ "acc": 0.2912912912912913
83
135
  },
84
136
  {
85
137
  "low": 0.3,
86
138
  "high": 0.4,
87
- "n": 5,
88
- "acc": 0.2
139
+ "n": 2416,
140
+ "acc": 0.3265728476821192
89
141
  },
90
142
  {
91
143
  "low": 0.4,
92
144
  "high": 0.5,
93
- "n": 9,
94
- "acc": 0.4444444444444444
145
+ "n": 4308,
146
+ "acc": 0.3152274837511606
95
147
  },
96
148
  {
97
149
  "low": 0.5,
98
150
  "high": 0.6,
99
- "n": 20,
100
- "acc": 0.4
151
+ "n": 4777,
152
+ "acc": 0.3382876282185472
101
153
  },
102
154
  {
103
155
  "low": 0.6,
104
156
  "high": 0.7,
105
- "n": 8,
106
- "acc": 0.5
157
+ "n": 4943,
158
+ "acc": 0.3568682986040866
107
159
  },
108
160
  {
109
161
  "low": 0.7,
110
162
  "high": 0.8,
111
- "n": 19,
112
- "acc": 0.3684210526315789
163
+ "n": 5534,
164
+ "acc": 0.382544271774485
113
165
  },
114
166
  {
115
167
  "low": 0.8,
116
168
  "high": 0.9,
117
- "n": 25,
118
- "acc": 0.4
169
+ "n": 8066,
170
+ "acc": 0.43627572526655095
119
171
  },
120
172
  {
121
173
  "low": 0.9,
122
174
  "high": 1.0,
123
- "n": 1114,
124
- "acc": 0.8824057450628366
175
+ "n": 30517,
176
+ "acc": 0.6440344725890488
125
177
  }
126
178
  ]
127
179
  },
@@ -130,7 +182,7 @@
130
182
  "particle-honorific kryptonite (e.g. FR 'Saint-Just-Saint-Rambert') if not in synth set",
131
183
  "non-Latin scripts (CJK, Cyrillic) fall through to byte-fallback tokens; F1 unknown"
132
184
  ],
133
- "notes": "Stage 1 coarse v0.2.0 \u2014 same architecture as v0.1.0 (8.87M params, 6L/256H/4-heads), trained on the expanded corpus-v0.2.0 (262.7M aligned rows, 6 train sources) with the loader rewrite from issue #43 (source-weighted multinomial sampler + relaxed coarse filter). The v0.1.0 positional-heuristic overfit was driven by a strict country-tag gate that dropped ~94% of v0.2.0 before any source weighting; with the gate relaxed and the loader interleaving sources at the row level, the model now sees a fixed mix of ban/tiger/nppes/state-tx/wof-admin/wof-postalcode per batch instead of mono-source blocks. See evals/scores-by-version.json for the v0.1.0 \u2192 v0.2.0 deltas.",
185
+ "notes": "v0.4.0 \u2014 issue #116. Same encoder geometry as v0.3.0 (8.87M params, 6L/256H/4-heads, 21 BIO labels, linear-chain CRF). Issue proposed (1) per-token CRF NLL normalization + (3) class-weighted CE biased toward coarse labels + (4) source-weight rebalance. Empirical iteration found that \u00a71 and \u00a73 destabilize sustained training at every LR tested (5e-4, 3e-4, 1.5e-4 \u2014 even the v0.3.0-safe LR), and on the golden v0.1.2 eval (4535 entries) they slightly REGRESS country/postcode F1 vs v0.3.0. SHIPPED recipe is the \u00a74-only ablation (v0.3.0 dual-loss + v0.4.0 source-weight rebalance) at lr=1.5e-4, step 2200. Modest fine-label gains (street +0.03, house_number +0.01), modest coarse-F1 regression (country -0.07, postcode -0.07). \u00a71/\u00a73 deferred to v0.4.1 corpus-side investigation per issue's '2K divergence' clause. Full ablation matrix retrospective in LOG.md.",
134
186
  "format": {
135
187
  "model": "ONNX int8 dynamic",
136
188
  "tokenizer": "SentencePiece unigram, byte_fallback=true, vocab_size=16000",
@@ -142,5 +194,5 @@
142
194
  "tokenizer": "tokenizer.model",
143
195
  "model_card": "model-card.json"
144
196
  },
145
- "base_relpath": "/data/models/checkpoints/stage1-coarse/step-050000"
197
+ "base_relpath": "/data/models/checkpoints/v0_4_0-stableLR-source-only/step-002200"
146
198
  }
package/model.onnx CHANGED
Binary file
package/package.json CHANGED
@@ -1,19 +1,19 @@
1
1
  {
2
- "name": "@mailwoman/neural-weights-fr-fr",
3
- "version": "2.0.0",
4
- "description": "Mailwoman neural-classifier weights for locale 'fr-fr'. Data-only package — loaded by @mailwoman/neural at runtime.",
5
- "license": "AGPL-3.0-only",
6
- "repository": {
7
- "type": "git",
8
- "url": "https://github.com/sister-software/mailwoman"
9
- },
10
- "files": [
11
- "model.onnx",
12
- "tokenizer.model",
13
- "model-card.json",
14
- "README.md"
15
- ],
16
- "publishConfig": {
17
- "access": "public"
18
- }
19
- }
2
+ "name": "@mailwoman/neural-weights-fr-fr",
3
+ "version": "2.2.0",
4
+ "description": "Mailwoman neural-classifier weights for locale 'fr-fr'. Data-only package — loaded by @mailwoman/neural at runtime.",
5
+ "license": "AGPL-3.0-only",
6
+ "repository": {
7
+ "type": "git",
8
+ "url": "https://github.com/sister-software/mailwoman"
9
+ },
10
+ "files": [
11
+ "model.onnx",
12
+ "tokenizer.model",
13
+ "model-card.json",
14
+ "README.md"
15
+ ],
16
+ "publishConfig": {
17
+ "access": "public"
18
+ }
19
+ }
package/tokenizer.model CHANGED
Binary file