license: agpl-3.0 language:

• en library_name: onnx pipeline_tag: token-classification tags:
• token-classification
• named-entity-recognition
• address-parsing
• postal-address
• geocoding
• onnx
• sequence-labeling

mailwoman — neural address-parser weights (en-nz)

The trained-model bundle (@mailwoman/neural-weights-en-nz) for Mailwoman, a postal-address parser. This package is data only — a model-card.json, the NZ placetype-pair retrieval index, and the shared gazetteer lexicons. It has no JavaScript logic of its own; it is loaded at inference time by @mailwoman/neural.

This locale is served by the shared multi-locale model. The en-nz bundle ships no model.onnx/tokenizer.model of its own — it declares @mailwoman/neural-weights-en-us as its mailwoman.baseWeights and resolves the base package's model + tokenizer at runtime (byte-identical artifact; one encoder serves both locales). What this package ships is the NZ-specific soft-feed data: the placetype-pair retrieval index (pair-index-nz.bin, built from the LINZ-derived OpenAddresses NZ countrywide register — see Evaluation below), plus the shared gazetteer/country lexicons.

What this is

The NZ-facing half of Mailwoman's retrieval side: real (child, parent) suburb/town pairs (e.g. "Plimmerton" is a real suburb of "Porirua") fed to the decoder as a soft dependent_locality bias, hard-gated to NZ input only. NZ's register genuinely repeats names across tiers — 21.6% of its pairs are identity pairs (suburb == town, e.g. "Mangawhai, Mangawhai") — which is why @mailwoman/neural ≥7.8.0's identical-adjacent-segment rule exists; this bundle's index is that rule's data source.

Unlike the en-gb sibling, no postcode-anchor binary ships — no WOF NZ postcode shard exists yet, so the postcode-anchor channel resolves OFF for en-nz (a loud one-time warning, not a crash). Building that shard is the tracked follow-up in model-card.json.

Intended use

Parsing free-text NZ postal addresses into structured components (locality, dependent_locality, street, house_number, …) for geocoding — resolving a parsed address to coordinates via a gazetteer/resolver.

Ship-config requirement (read before using)

The Mailwoman model expects the soft anchor + gazetteer channels fed at inference. Construct the scorer through @mailwoman/neural's createScorer (the canonical ProductionScorer), which reads the bundle's requires/channel contract and fails closed if a declared channel isn't fed. Do not hand-wire the raw ONNX session with the anchor input zero-filled.

Evaluation

en-nz battery, 2026-07-24 — all 6 pre-registered bars PASS. Prior ON at the calibrated δ=10, on the shipped v385 base: NZ suburb board (246 rows) as-written 246/246 emission, 246/246 tag-correct (100%); comma-stripped 244/246 (99.2%); curated no-suburb board 0/54 false positives. The venue-confound read is interim — 0/510 on a synthetic board; no real NZ venue-name source exists on disk yet (issue #1279). Every GB number reproduces exactly through the same code path. Full breakdown, δ-sweep table, and the repeated-name-convention story: model-card.json's eval and notes blocks.

Limitations

• Expects its channels — see Ship-config requirement.
• No NZ postcode anchor yet — the anchor channel is OFF for en-nz until a WOF NZ postcode shard is built (model-card follow-up).
• Venue-confound specificity is interim — measured 0 FP on a synthetic board only; real NZ venue-name data is an open acquisition (issue #1279).
• Placetype-pair prior is hard-gated to NZ — it structurally cannot fire on non-NZ input.
• All-caps / shouting input degrades the admin tags (mixed-case training); @mailwoman/neural's normalizeCase opt recovers detected all-caps ASCII.
• Non-Latin scripts (CJK, Cyrillic) fall through to byte-fallback tokens; quality there is unmeasured.

License & links

• License: AGPL-3.0-only.
• Docs & concepts: https://mailwoman.sister.software
• Loader / scorer API: @mailwoman/neural (createScorer).
• The functional contract for this bundle is model-card.json; this README.md is the HuggingFace-facing card.