@mailwoman/neural-weights-fr-fr 2.0.0

package/README.md

@@ -0,0 +1,37 @@
+# @mailwoman/neural-weights-fr-fr
+
+Phase 2 / Stage 1 (coarse) Mailwoman neural-classifier weights.
+
+- locale: **fr-fr**
+- corpus: **0.2.0**
+- training steps: **50000**
+- hardware: **AMD Radeon 780M (gfx1103) bf16 ~14.6 GiB GTT**
+
+## Phase 2 §6 status
+
+**⚠ Below Phase 2 §6 targets (≥95% F1):**
+
+- `country` F1 = **0.0000** (target ≥0.95)
+- `region` F1 = **0.8293** (target ≥0.95)
+- `locality` F1 = **0.6471** (target ≥0.95)
+- `postcode` F1 = **0.8594** (target ≥0.95)
+
+## Eval (golden set)
+
+- entries: **74**
+- full-parse exact match: **0.5270**
+- mean token confidence: **0.9745**
+
+## Components supported
+
+Stage 1 ships coarse-only: country / region / locality / dependent_locality / postcode / subregion / cedex. Street- and venue-level components are explicit future phases.
+
+## Files
+
+- `model.onnx` — int8-quantized ONNX model.
+- `tokenizer.model` — SentencePiece unigram tokenizer (matches the corpus version).
+- `model-card.json` — ModelCard with training + eval metadata.
+
+## Loader
+
+Loaded at runtime by `@mailwoman/neural`. This package contains no JS code.

package/model-card.json

@@ -0,0 +1,146 @@
+{
+	"name": "neural-weights-fr-fr",
+	"version": "0.2.0",
+	"phase": "Stage 1 (coarse)",
+	"license": "AGPL-3.0-only",
+	"locale": "fr-fr",
+	"training": {
+		"corpus_version": "0.2.0",
+		"tokenizer_version": "0.1.0",
+		"steps": 50000,
+		"hardware": "AMD Radeon 780M (gfx1103) bf16 ~14.6 GiB GTT",
+		"duration_seconds": 23520.0,
+		"started_at": null,
+		"completed_at": "2026-05-18T21:33:27.380730Z"
+	},
+	"components_supported": ["country", "region", "locality", "dependent_locality", "postcode", "subregion", "cedex"],
+	"eval": {
+		"n_entries": 74,
+		"full_parse_exact_match": 0.527027027027027,
+		"mean_token_confidence": 0.974534777700901,
+		"per_component": {
+			"country": {
+				"precision": 0.0,
+				"recall": 0.0,
+				"f1": 0.0,
+				"support": 6
+			},
+			"region": {
+				"precision": 0.8499999999858334,
+				"recall": 0.80952380951096,
+				"f1": 0.8292682921697403,
+				"support": 63
+			},
+			"locality": {
+				"precision": 0.6874999999892578,
+				"recall": 0.6111111111026234,
+				"f1": 0.6470588230216262,
+				"support": 72
+			},
+			"dependent_locality": {
+				"precision": 0.0,
+				"recall": 0.0,
+				"f1": 0.0,
+				"support": 1
+			},
+			"postcode": {
+				"precision": 0.8730158730020157,
+				"recall": 0.8461538461408283,
+				"f1": 0.8593749994866943,
+				"support": 65
+			},
+			"subregion": {
+				"precision": 0.0,
+				"recall": 0.0,
+				"f1": 0.0,
+				"support": 0
+			},
+			"cedex": {
+				"precision": 0.0,
+				"recall": 0.0,
+				"f1": 0.0,
+				"support": 1
+			}
+		},
+		"calibration": [
+			{
+				"low": 0.0,
+				"high": 0.1,
+				"n": 0,
+				"acc": 0.0
+			},
+			{
+				"low": 0.1,
+				"high": 0.2,
+				"n": 0,
+				"acc": 0.0
+			},
+			{
+				"low": 0.2,
+				"high": 0.3,
+				"n": 0,
+				"acc": 0.0
+			},
+			{
+				"low": 0.3,
+				"high": 0.4,
+				"n": 5,
+				"acc": 0.2
+			},
+			{
+				"low": 0.4,
+				"high": 0.5,
+				"n": 9,
+				"acc": 0.4444444444444444
+			},
+			{
+				"low": 0.5,
+				"high": 0.6,
+				"n": 20,
+				"acc": 0.4
+			},
+			{
+				"low": 0.6,
+				"high": 0.7,
+				"n": 8,
+				"acc": 0.5
+			},
+			{
+				"low": 0.7,
+				"high": 0.8,
+				"n": 19,
+				"acc": 0.3684210526315789
+			},
+			{
+				"low": 0.8,
+				"high": 0.9,
+				"n": 25,
+				"acc": 0.4
+			},
+			{
+				"low": 0.9,
+				"high": 1.0,
+				"n": 1114,
+				"acc": 0.8824057450628366
+			}
+		]
+	},
+	"known_failure_modes": [
+		"underperforms on Hawaiian addresses (sparse in training corpus)",
+		"particle-honorific kryptonite (e.g. FR 'Saint-Just-Saint-Rambert') if not in synth set",
+		"non-Latin scripts (CJK, Cyrillic) fall through to byte-fallback tokens; F1 unknown"
+	],
+	"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.",
+	"format": {
+		"model": "ONNX int8 dynamic",
+		"tokenizer": "SentencePiece unigram, byte_fallback=true, vocab_size=16000",
+		"max_sequence_length": 128,
+		"opset": 17
+	},
+	"files": {
+		"model": "model.onnx",
+		"tokenizer": "tokenizer.model",
+		"model_card": "model-card.json"
+	},
+	"base_relpath": "/data/models/checkpoints/stage1-coarse/step-050000"
+}

package/package.json

@@ -0,0 +1,19 @@
+{
+ "name": "@mailwoman/neural-weights-fr-fr",
+ "version": "2.0.0",
+ "description": "Mailwoman neural-classifier weights for locale 'fr-fr'. Data-only package — loaded by @mailwoman/neural at runtime.",
+ "license": "AGPL-3.0-only",
+ "repository": {
+  "type": "git",
+  "url": "https://github.com/sister-software/mailwoman"
+ },
+ "files": [
+  "model.onnx",
+  "tokenizer.model",
+  "model-card.json",
+  "README.md"
+ ],
+ "publishConfig": {
+  "access": "public"
+ }
+}