@mailwoman/neural-weights-cjk

Mailwoman neural-classifier weights for CJK scripts: the char-path base model, Japanese, Korean, Taiwanese and Chinese under one 49-label head (stage3-cjk). Data-only; @mailwoman/neural loads it at runtime.

The manifest lists model.onnx, which mwops release copy-weights materializes from release.config.json's charWeights.cjk at release time; the ja-jp / zh-cn overlays are data-only packages over this base.

What this package ships

• model.onnx — the char graph: inputs char_ids int64 (batch, sequence, 7) and attention_mask int64 (batch, sequence), output logits (batch, sequence, 49). No input_ids.
• char-vocab.json — the sealed character vocabulary (4,451 entries, <pad> 0, <unk> 1, code-point order). This is the model's whole tokenizer: one unit per Unicode code point, a ±3 character window per unit, 96 units per row.
• model-card.json — the encoder: "char" block (char_vocab, max_units, max_unit_width, char_ctx), the 49 BIO labels, the training provenance, and the board reads.

What this package does NOT ship

• No tokenizer.model. A char graph has no SentencePiece vocabulary; weights resolution requires char-vocab.json in its place (packages/neural/lib/weights/index.ts, the card's encoder block).
• No soft-feed channels. The char path is channel-free by contract: no postcode-anchor, gazetteer, country or evidence lexicons, and the graph declares no channel inputs.
• No FST autocomplete artifact. fst-ja-jp.bin / fst-zh-cn.bin ship in the @mailwoman/neural-weights-ja-jp and -zh-cn overlays, data-only packages over this base.

Loading it today

import { NeuralAddressClassifier } from "@mailwoman/neural/classifier"

const classifier = await NeuralAddressClassifier.loadFromWeights({
	modelPath: "<package>/model.onnx",
	charVocabPath: "<package>/char-vocab.json",
	modelCardPath: "<package>/model-card.json",
})

await classifier.parseJSON("東京都千代田区丸の内1丁目9-1")
// { prefecture: "東京都", municipality: "千代田区", district: "丸の内", block: "1丁目", house_number: "9-1" }

Provenance

v8-cjk-regs seed 42, 8,000 steps from scratch on the v8-jp-kana JP corpus (2,000,000 rows, Overture-JP, five registers including the municipality's kana reading), the Korean road-name address register rebuilt from the ministry's 2026-08 주소DB (2,000,000 rows, seven registers including the lot-number form), the Taiwan corpus from Overture-TW (2,000,000 rows, five registers), three registries of typed business addresses aligned exactly against those keys before training (Korean permits, Taiwanese companies, Japanese corporate numbers; 1,666,000 rows), and 126 Chinese organizational-unit rows. On the 20,000-row held-out JP board the native register reads 0.9954 acceptability at 15 km against the previous base's 0.9921 on the same scorer, and かすみがうら市 fails 0 of 823 rows. On the 20,000-row Korean board over 26 held-out 시군구 the spans read region 1.000 / subregion 0.998 / dependent_locality 0.998 / street 0.999 / house_number 1.000 / postcode 1.000; on the 20,000-row Taiwan board over 28 held-out 鄉鎮市區, region 1.000 / subregion 1.000 / street 0.999 / house_number 0.999. The 町 whose names carry 市 are closed at decode time by the register in @mailwoman/codex/jp (#2178). Decision record: docs/superpowers/specs/2026-09-05-cjk-serving-path.md; the run record is docs/records/evals/2026-09-08-v8-cjk-regs.md; receipts on #1176, #2034, #2164, #2184 and #2204.

Dev setup

The graph is not committed. Link it into the data-root overlay from release.config.json's charWeights.cjk:

node packages/neural-weights-cjk/scripts/link-dev-weights.ts

At release time mwops release copy-weights materializes the same two files into the workspace from that recipe.