@mailwoman/neural-weights-fr-fr 4.1.0 → 4.3.0

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 +3 -3
  2. package/model.onnx +0 -0
  3. package/package.json +1 -1
package/model-card.json CHANGED
@@ -1,7 +1,7 @@
1
1
  {
2
2
  "name": "neural-weights-fr-fr",
3
- "version": "4.1.0",
4
- "model_lineage": "shares the en-us v0.9.7-unit-v3 multi-locale model (step 20000) — shipped as the unified 4.1.0 release version; tokenizer 0.6.0-a0. fr-fr and en-us bundle the SAME byte-identical model; this package carries the FR-specific calibration table.",
3
+ "version": "4.3.0",
4
+ "model_lineage": "shares the en-us v1.0.2-consolidation-runB multi-locale model (step 20000) — shipped as the unified 4.2.0 release version; FR postcode 99.6 / house_number 94.6 (best ever) / region 27.6 (open gap #330)",
5
5
  "phase": "Stage 3 — multi-locale (FR via the shared model)",
6
6
  "license": "AGPL-3.0-only",
7
7
  "locale": "fr-fr",
@@ -181,7 +181,7 @@
181
181
  "particle-honorific kryptonite (e.g. FR 'Saint-Just-Saint-Rambert') if not in synth set",
182
182
  "non-Latin scripts (CJK, Cyrillic) fall through to byte-fallback tokens; F1 unknown"
183
183
  ],
184
- "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.",
184
+ "notes": "v0.4.0 — 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 §1 and §3 destabilize sustained training at every LR tested (5e-4, 3e-4, 1.5e-4 — 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 §4-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). §1/§3 deferred to v0.4.1 corpus-side investigation per issue's '2K divergence' clause. Full ablation matrix retrospective in LOG.md.",
185
185
  "format": {
186
186
  "model": "ONNX int8 dynamic",
187
187
  "tokenizer": "SentencePiece unigram, byte_fallback=true, vocab_size=16000",
package/model.onnx CHANGED
Binary file
package/package.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "@mailwoman/neural-weights-fr-fr",
3
- "version": "4.1.0",
3
+ "version": "4.3.0",
4
4
  "description": "Mailwoman neural-classifier weights for locale 'fr-fr'. Data-only package — loaded by @mailwoman/neural at runtime.",
5
5
  "license": "AGPL-3.0-only",
6
6
  "repository": {