Context

What Mailwoman is today

Mailwoman is a fork of pelias/parser, written in TypeScript. It is a rule-based natural language classification engine for geocodinggeocodingThe process of converting an address into geographic coordinates (latitude and longitude). Mailwoman geocodes in a multi-tier cascade: exact address-point match → street interpolation → locality centroid. Each tier is progressively coarser but more widely available.. Input: an address string. Output: a list of solutions, each a set of { component, confidence, offset, penalty } spansspanA contiguous range of characters or tokens in the input string, tagged with an address component type (street, locality, postcode, etc.). Parsed addresses are represented as collections of spans, possibly nested in a tree..

Architecture: tokenization (sections → words → phrases) → classification (dictionary/regex/composite classifiers populate a graph) → solving (ExclusiveCartesianSolver enumerates valid permutations, additional solvers filter and augment).

Output example for "Mt Tabor Park, 6220 SE Salmon St, Portland, OR 97215, USA":

;[
	{ venue: "Mt Tabor Park", confidence: 0.8, offset: 0, penalty: 0 },
	{ house_number: "6220", confidence: 0.9, offset: 15, penalty: 0 },
	{ street: "SE Salmon St", confidence: 0.98, offset: 20, penalty: 0 },
	{ locality: "Portland", confidence: 1, offset: 34, penalty: 0 },
	{ region: "OR", confidence: 1, offset: 44, penalty: 0 },
	{ postcode: "97215", confidence: 1, offset: 47, penalty: 0 },
	{ country: "USA", confidence: 0.9, offset: 54, penalty: 0 },
]

The shape — per-component confidence + character offset + penalty — is unusually good for ML integration. A tokentokenOne word or subword in the tokenized input. For the neural classifier, tokens come from SentencePiece (subword units); for the rule classifiers, tokens are whitespace- and punctuation-separated words.-classification modelneural classifierThe machine learning model at the core of Mailwoman's parser — a transformer encoder (~30M parameters) trained from scratch to do BIO token classification over addresses. It learns the 'grammar' of address formats; the gazetteer supplies the 'atlas.''s natural output is essentially this. We are not adapting a modelneural classifierThe machine learning model at the core of Mailwoman's parser — a transformer encoder (~30M parameters) trained from scratch to do BIO token classification over addresses. It learns the 'grammar' of address formats; the gazetteer supplies the 'atlas.' to fit a foreign output shape.

What the field looks like

• libpostallibpostalAn open-source C address parser used by Pelias. Mailwoman's rule-based v0 and neural classifier supersede it. (C, CRFCRF (Conditional Random Field). A statistical modeling method that predicts structured outputs by modeling dependencies between adjacent labels. Mailwoman uses a linear-chain CRF as the Viterbi decoder at inference time to enforce BIO label consistency — a B-street must be followed by I-street or O, never I-locality.) — the de-facto baseline. 99.45% claimed parseaddress parsingThe process of decomposing a free-text postal address string into structured components — house number, street name, locality, region, postcode, and country — so a geocoder can resolve them to coordinates. accuracy on held-out data. Trained on OpenStreetMapOpenStreetMap (OSM). A community-curated global map database (ODbL-licensed) with addr:* tagged features and place hierarchies. A secondary corpus source and a source of street names.. Mostly stagnant since ~2018.
• Deepparse (Python, Seq2Seq BiLSTM) — academic, fine-tunable, slower than libpostallibpostalAn open-source C address parser used by Pelias. Mailwoman's rule-based v0 and neural classifier supersede it., >99% component accuracy. Active.
• PeliasPeliasAn open-source geocoder, Mailwoman's spiritual predecessor. Parser (JS, rule-based) — our upstream. Explicitly rejects black-box ML in favor of debuggable rules. Active.
• Airmail (Rust, tantivy-based geocoder) — sidesteps parsingaddress parsingThe process of decomposing a free-text postal address string into structured components — house number, street name, locality, region, postcode, and country — so a geocoder can resolve them to coordinates. entirely by treating it as an IR problem. 376 stars, single-maintainer, planet-scale for ~$5/month. Worth studying.
• Transformerneural classifierThe machine learning model at the core of Mailwoman's parser — a transformer encoder (~30M parameters) trained from scratch to do BIO token classification over addresses. It learns the 'grammar' of address formats; the gazetteer supplies the 'atlas.'-based parsers in academic literature — Guermazi et al. 2024 shows transformersneural classifierThe machine learning model at the core of Mailwoman's parser — a transformer encoder (~30M parameters) trained from scratch to do BIO token classification over addresses. It learns the 'grammar' of address formats; the gazetteer supplies the 'atlas.' beat libpostallibpostalAn open-source C address parser used by Pelias. Mailwoman's rule-based v0 and neural classifier supersede it. on noisy/multilingual data. Not productized.

The gap: there is no widely adopted neural address parser shipped as a library. Everyone either uses 2018-era CRFCRF (Conditional Random Field). A statistical modeling method that predicts structured outputs by modeling dependencies between adjacent labels. Mailwoman uses a linear-chain CRF as the Viterbi decoder at inference time to enforce BIO label consistency — a B-street must be followed by I-street or O, never I-locality. (libpostallibpostalAn open-source C address parser used by Pelias. Mailwoman's rule-based v0 and neural classifier supersede it.) or rolls their own.

Why neural, why now

Three reasons rules hit a ceiling:

1. Ambiguity that needs context. "Paris Texas" vs "Paris, Texas" vs "Paris, France." Rules need explicit comma logic; a modelneural classifierThe machine learning model at the core of Mailwoman's parser — a transformer encoder (~30M parameters) trained from scratch to do BIO token classification over addresses. It learns the 'grammar' of address formats; the gazetteer supplies the 'atlas.' learns context.
2. Internationalization without rule explosion. Every countrycountryThe top-level address component (an ISO country). Closed-vocabulary, so it is best handled by a deterministic matcher feeding a proposal rather than a retrained model head.'s rule set is bespoke maintenance. A modelneural classifierThe machine learning model at the core of Mailwoman's parser — a transformer encoder (~30M parameters) trained from scratch to do BIO token classification over addresses. It learns the 'grammar' of address formats; the gazetteer supplies the 'atlas.' trained on countrycountryThe top-level address component (an ISO country). Closed-vocabulary, so it is best handled by a deterministic matcher feeding a proposal rather than a retrained model head.-tagged data generalizes.
3. Graceful degradation on novel input. Rules either match or don't. A modelneural classifierThe machine learning model at the core of Mailwoman's parser — a transformer encoder (~30M parameters) trained from scratch to do BIO token classification over addresses. It learns the 'grammar' of address formats; the gazetteer supplies the 'atlas.' produces probabilities, enabling confidence-weighted decisions in the solver.

But rules win in three places:

1. PostcodespostcodeThe country-specific postal code (US ZIP, French code postal, etc.). Mailwoman handles postcode parsing entirely by rule classifier — a regex problem, not an ML one.. A US ZIP is a regex. A modelneural classifierThe machine learning model at the core of Mailwoman's parser — a transformer encoder (~30M parameters) trained from scratch to do BIO token classification over addresses. It learns the 'grammar' of address formats; the gazetteer supplies the 'atlas.' is wasted capacity.
2. Explicit dictionaries. CountrycountryThe top-level address component (an ISO country). Closed-vocabulary, so it is best handled by a deterministic matcher feeding a proposal rather than a retrained model head. names, stateregionThe first-level administrative subdivision of a country — a US state, a French region, a province. The component between country and locality. abbreviations — lookups are perfect and instant.
3. Debuggability. When a rule misclassifies, you can read the rule. When a modelneural classifierThe machine learning model at the core of Mailwoman's parser — a transformer encoder (~30M parameters) trained from scratch to do BIO token classification over addresses. It learns the 'grammar' of address formats; the gazetteer supplies the 'atlas.' misclassifies, you train on more data and hope.

Hence Ship of TheseusShip of TheseusThe migration pattern Mailwoman uses: replace rule classifiers with the neural classifier one component at a time, only when metrics justify it. Named for the philosophical thought experiment.: migrate component-by-component, gated on metrics, never flip a single "neural on" switch.

Why TypeScript-first

Mailwoman's users are in the Node ecosystem. A Python or Rust dependency is a deployment regression for them. ONNX RuntimeONNX (Open Neural Network Exchange). An open format for machine learning models that enables interoperability between training frameworks and inference runtimes. Mailwoman ships its trained model as an ONNX file so it can run in Node.js and the browser via onnxruntime. has a stable Node binding (onnxruntime-node) with CPU + CUDA execution providers. SentencePieceSentencePieceA language-independent subword tokenizer that splits text into pieces using a unigram language model. Mailwoman uses a SentencePiece tokenizer with a 48,000-token vocabulary and byte-fallback, trained on address data rather than general text. has WASM builds. The full inferenceinferenceRunning the trained model on new input to get predictions, as opposed to training, which produces the model. In Mailwoman that means a small transformer encoder reads an address string and classifies every token — house number, street, locality, region, postcode, and the rest. A Who's On First gazetteer can feed soft location hints into the pass, but the model makes the final call on every label. Where a generative model writes text token by token, Mailwoman's output is a retrieval-augmented token classification: one label per input piece. path runs in a Node process without spawning subprocesses or shelling out.

TrainingtrainingThe process of adjusting a model's parameters so its predictions match labeled examples, by repeatedly measuring error and nudging the weights to reduce it. Distinct from inference, when the trained model is run on new input. stays in Python because the ecosystem (HuggingFace Transformers, PyTorch, the datasets library, ONNXONNX (Open Neural Network Exchange). An open format for machine learning models that enables interoperability between training frameworks and inference runtimes. Mailwoman ships its trained model as an ONNX file so it can run in Node.js and the browser via onnxruntime. export tooling) is unmatched. The trainingtrainingThe process of adjusting a model's parameters so its predictions match labeled examples, by repeatedly measuring error and nudging the weights to reduce it. Distinct from inference, when the trained model is run on new input. code is internal to the project and never published to npm.

Why US + France first

• US: the maintainer's home market, rule classifiers are already strong here, easiest to A/B against.
• France: non-Anglo, non-trivial grammar (particles de la/du/des, CEDEXCEDEX (Courrier d'Entreprise à Distribution Exceptionnelle). A French postal routing for high-volume business mail: a CEDEX code delivers directly from a sorting centre, bypassing the local post office. A common negative-space format Mailwoman must parse. postal routing, arrondissement notation), official open data (BANBAN (Base Adresse Nationale). France's authoritative open national address register — the highest-quality training source for French addresses, with full component structure.) that's first-class, and uses Latin script so we don't yet need to solve tokenizertokenizerThe component that converts a raw address string into a sequence of numeric token IDs the model can process. Mailwoman's tokenizer is a SentencePiece unigram model trained specifically on postal addresses. vocabularyvocabularyThe fixed set of tokens a tokenizer can produce. Mailwoman's SentencePiece vocabulary is tens of thousands of subword pieces, with byte fallback for anything outside it. problemsexpectation-maximizationAn iterative algorithm that estimates model parameters when some variables are unobserved. In Mailwoman's matcher, EM learns the Fellegi-Sunter m and u parameters from unlabeled data — no training labels needed..

If the architecture handles FR cleanly, it will handle DE/IT/ES/PT trivially.

Why Japan as the validation milestone

Japanese addressing is structurally different from Western addressing: no streetsstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels., nested blocks (chōmechōmeIn Japanese addressing, a district-level subdivision in the block-based chōme / banchi / gō numbering scheme, which uses area-and-block numbers instead of street names./banchibanchiIn Japanese addressing, a block within a chōme. The middle level of the chōme / banchi / gō hierarchy./go), written right-to-fine-grained. If the schema, classifier interface, and policy system survive JP without core refactor, the architecture is sound. If they don't, we learn it in PhasephaseA milestone in the implementation plan (Foundation, Corpus, Training, Integration, and forward-looking phases). Distinct from stage (runtime pipeline) and tier (model vocabulary). 6 and not in production five years from now.

The bitter lesson, applied to address parsing

The project's design intent originated in Richard Sutton's "Bitter Lessonbitter lessonSutton's observation that general methods which scale with compute and data ultimately beat hand-engineered systems. The rationale behind Mailwoman's Ship-of-Theseus migration from rule classifiers to a learned model." essay. The thesis applied here: rule-based classifiersrule-based classifierMailwoman's legacy v0 parser — a library of deterministic token classifiers (house number, street suffix, postcode, place name, etc.) composed by priority. Now primarily used for corpus labeling, fallback classification, and arbitration diagnostics. + Cartesian solvers approximate a human's address-parsingaddress parsingThe process of decomposing a free-text postal address string into structured components — house number, street name, locality, region, postcode, and country — so a geocoder can resolve them to coordinates. intuition with hand-crafted machinery; given sufficient corpuscorpusThe BIO-labeled training data used to train Mailwoman's neural classifier. Assembled from real sources (OpenAddresses, National Address Database) and synthetic shards (boundary stress, order variants, negative space). Managed by @mailwoman/corpus., a language modellanguage model (LM). A model that assigns probabilities to sequences of tokens. Used here mostly as a prior — an FST or n-gram model that biases the decoder toward plausible sequences via shallow fusion. approximates that intuition directly — and outperforms — without the scaffolding.

The canonical short-form, from the operator's "Paris, Texas" talk (PeerTube: peertube.openstreetmap.fr video e776d210-f857-4df2-83a7-4414fd52f6f2, slide 19):

A geocoder is a contextual parser + constraint solver. Not a tokenizertokenizerThe component that converts a raw address string into a sequence of numeric token IDs the model can process. Mailwoman's tokenizer is a SentencePiece unigram model trained specifically on postal addresses.. Not a dumb database lookup.

The neural modelneural classifierThe machine learning model at the core of Mailwoman's parser — a transformer encoder (~30M parameters) trained from scratch to do BIO token classification over addresses. It learns the 'grammar' of address formats; the gazetteer supplies the 'atlas.' is the contextual-parser part. The constraint solver still has a role — it weighs which hierarchy interpretation is consistent (see the Arc-de-Triomphe disambiguation in slides 12→13 of that talk). The "dumb database lookup" pattern is what's being deprecated.

The current ExclusiveCartesianSolver inherited from PeliasPeliasAn open-source geocoder, Mailwoman's spiritual predecessor. is what slide 17 calls "Sparkling Bogosort":

"It's only 'Cartesian permutation' if your gazetteergazetteerA geographical index that maps place names and postcodes to real-world coordinates. Mailwoman uses a custom-built Who's On First (WOF) SQLite database as its gazetteer — the 'atlas' half of the grammar/atlas architecture. comes from France. Otherwise it's just 'Sparkling Bogosort'."

France's BANBAN (Base Adresse Nationale). France's authoritative open national address register — the highest-quality training source for French addresses, with full component structure. gives enough constraint that the permutation reduces to a small set; everywhere else it's randomly trying combinations and squinting at which one looks valid. The PolicyRegistry (see reference/INTERFACES.md) is the mechanism by which this gets phased out, component-by-component — the Ship-of-Theseus migration referenced above.

Adversarial / kryptonite cases the corpus must include

Slide 14 of the talk's "programming noir" debug list, plus operator additions:

• Buffalo Health Clinic Buffalo, New York — venuevenueA named, non-address place — a business, building, park, or stadium. Mailwoman's free-text point-of-interest component, added as a Tier 2 fine label. name overlaps with localitylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighbourhood in the hierarchy. tokentokenOne word or subword in the tokenized input. For the neural classifier, tokens come from SentencePiece (subword units); for the rule classifiers, tokens are whitespace- and punctuation-separated words.
• New York, New York Steakhouse, Las Vegas, Nevada — venuevenueA named, non-address place — a business, building, park, or stadium. Mailwoman's free-text point-of-interest component, added as a Tier 2 fine label. contains a place-shaped sub-string, then a real address follows
• Paris Texas vs Paris, Texas vs Paris, France — punctuation + context disambiguates
• St. Petersburg, Russia — St. is "Saint" not "StreetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels."
• P'tit St. Denis' Street Café — apostrophe-contracted French (P'tit for Petit) layered onto the St./Saint/StreetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels. ambiguity
• Saint-Denis vs Saint Denis vs St. Denis — same place, three orthographies; the modelneural classifierThe machine learning model at the core of Mailwoman's parser — a transformer encoder (~30M parameters) trained from scratch to do BIO token classification over addresses. It learns the 'grammar' of address formats; the gazetteer supplies the 'atlas.' must learn the alternation, not pick a canonical and reject the rest

These are not noise. They are the inputs a modelneural classifierThe machine learning model at the core of Mailwoman's parser — a transformer encoder (~30M parameters) trained from scratch to do BIO token classification over addresses. It learns the 'grammar' of address formats; the gazetteer supplies the 'atlas.' must learn to handle, and the inputs a rule parser cannot. The corpuscorpusThe BIO-labeled training data used to train Mailwoman's neural classifier. Assembled from real sources (OpenAddresses, National Address Database) and synthetic shards (boundary stress, order variants, negative space). Managed by @mailwoman/corpus. pipelinestaged pipelineMailwoman's runtime architecture: a sequence of pure-function stages (normalize → query-shape → locale-gate → kind-classifier → phrase-grouper → classifier → decoder) connected by typed handoffs. Each stage is published as its own npm package. that produces trainingtrainingThe process of adjusting a model's parameters so its predictions match labeled examples, by repeatedly measuring error and nudging the weights to reduce it. Distinct from inference, when the trained model is run on new input. data must include them — generated alongside their correct labelscomponent tagOne of the 33 labels in Mailwoman's address schema — street, locality, region, postcode, house_number, unit, po_box, country, venue, intersection, and others. Each parsed span carries exactly one component tag. (compositional synthesis), not relied on the alignmentalignmentThe step in the corpus pipeline that takes a (raw, components) pair from an adapter and produces a (raw, tokens, BIO labels) row by finding each component's text inside the raw string and labeling the matching tokens. layerlayerOne transformer block — attention plus a feed-forward network, with normalization and residual connections — applied to every position. Stacking layers lets the model build up richer representations; Mailwoman's encoder has 6. to discover them after the fact.

Success metric: graceful failure

From slide 20 of the same talk:

People just want what their local post office expects. And, that means failing gracefully.

This is the deliberate counterpoint to F1-maximization. The mail carrier squinting at bad handwriting and figuring it out anyway — that is the target behavior. A wrong answer that at least looks plausible and routes to a recoverable next step is preferable to wrong-with-high-confidence.

The confidence-as-trap pattern

A core operating principle for everyone making decisions about confidence thresholds, scoring tweaks, or post-hoc adjustments in the parser:

Historically the confidence levels and various phrase constraints, confidence boosts, and demerits in rule-classifier-era geocoders are all part of a greater problem. As one builds more and more exceptions and adjustments to a domain-specific machine, one occludes the true nature of how addresses actually work. And worse, one's mental confidence in how accurately the domain has been mapped reinforces the belief — not with a platonic ideal of well-formed addresses, and not even of the platonic ideal of what a badly formatted address might look like, but with only the addresses which the system happens to struggle with the most. At the beginning, luck is in one's favor and meaningful exceptions surface naturally. But each bug report of a badly parsed address entrenches the system deeper. And if one then tries a new way of parsingaddress parsingThe process of decomposing a free-text postal address string into structured components — house number, street name, locality, region, postcode, and country — so a geocoder can resolve them to coordinates., regressions are guaranteed — meaning even more machinery is required.

The principle that inverts this trap, and that should be quoted verbatim wherever this section is referenced:

We want the opposite: every additional address should give us confidence in the possibilities, not the constraints.

Operational implications

• PhasephaseA milestone in the implementation plan (Foundation, Corpus, Training, Integration, and forward-looking phases). Distinct from stage (runtime pipeline) and tier (model vocabulary). 2 trainingtrainingThe process of adjusting a model's parameters so its predictions match labeled examples, by repeatedly measuring error and nudging the weights to reduce it. Distinct from inference, when the trained model is run on new input. metrics must capture confidence calibrationconfidence calibrationThe process of adjusting model confidence scores so that '0.6' actually means the model is right about 60% of the time. Mailwoman uses isotonic regression (PAVA) to calibrate per-span confidences against held-out data. Applied opt-in via createCalibrator. and graceful-degradation rate, not just per-component F1. A modelneural classifierThe machine learning model at the core of Mailwoman's parser — a transformer encoder (~30M parameters) trained from scratch to do BIO token classification over addresses. It learns the 'grammar' of address formats; the gazetteer supplies the 'atlas.' that scores well on F1 while reporting overconfident wrong answers on ambiguous inputs fails the success metric.
• ClassifierPolicy.confidence_threshold tuning posture: pick values that prefer "no answer + ask" over "wrong with high confidence." Never tune against specific user-reported failures — that's the entrenchment cycle starting.
• Bug-triage process: when a parser bug is reported, add a trainingtrainingThe process of adjusting a model's parameters so its predictions match labeled examples, by repeatedly measuring error and nudging the weights to reduce it. Distinct from inference, when the trained model is run on new input. row, not a rule or a confidence demerit. The bug report becomes a corpuscorpusThe BIO-labeled training data used to train Mailwoman's neural classifier. Assembled from real sources (OpenAddresses, National Address Database) and synthetic shards (boundary stress, order variants, negative space). Managed by @mailwoman/corpus. entry; the modelneural classifierThe machine learning model at the core of Mailwoman's parser — a transformer encoder (~30M parameters) trained from scratch to do BIO token classification over addresses. It learns the 'grammar' of address formats; the gazetteer supplies the 'atlas.' learns from the example without anyone touching the parser machinery. If a bug surfaces a missing schema tag, that's a PhasephaseA milestone in the implementation plan (Foundation, Corpus, Training, Integration, and forward-looking phases). Distinct from stage (runtime pipeline) and tier (model vocabulary). 0 revisit (see reference/SCHEMA.md), not a confidence patch.
• Golden evalevalRunning the model against a held-out golden dataset and computing per-component F1, exact-match, calibration, and resolved-coordinate error. set must include intentionally-ambiguous addresses where the correct answer is "partial parseaddress parsingThe process of decomposing a free-text postal address string into structured components — house number, street name, locality, region, postcode, and country — so a geocoder can resolve them to coordinates. + low-confidence flag" — not just the unambiguous cases.
• No confidence_boost / confidence_demerit layerlayerOne transformer block — attention plus a feed-forward network, with normalization and residual connections — applied to every position. Stacking layers lets the model build up richer representations; Mailwoman's encoder has 6. in the synthesis pipelinestaged pipelineMailwoman's runtime architecture: a sequence of pure-function stages (normalize → query-shape → locale-gate → kind-classifier → phrase-grouper → classifier → decoder) connected by typed handoffs. Each stage is published as its own npm package. or downstream of the modelneural classifierThe machine learning model at the core of Mailwoman's parser — a transformer encoder (~30M parameters) trained from scratch to do BIO token classification over addresses. It learns the 'grammar' of address formats; the gazetteer supplies the 'atlas.'. That's the rule-era machinery the bitter lessonbitter lessonSutton's observation that general methods which scale with compute and data ultimately beat hand-engineered systems. The rationale behind Mailwoman's Ship-of-Theseus migration from rule classifiers to a learned model. rejects. The modelneural classifierThe machine learning model at the core of Mailwoman's parser — a transformer encoder (~30M parameters) trained from scratch to do BIO token classification over addresses. It learns the 'grammar' of address formats; the gazetteer supplies the 'atlas.''s softmaxsoftmaxThe function that converts a vector of logits into a probability distribution summing to 1, applied after priors and biases are added to the emission logits. probability is the confidence; calibration is a trainingtrainingThe process of adjusting a model's parameters so its predictions match labeled examples, by repeatedly measuring error and nudging the weights to reduce it. Distinct from inference, when the trained model is run on new input.-time concern.

What this project is not

• Not a geocoder. Resolution to coordinates is PhasephaseA milestone in the implementation plan (Foundation, Corpus, Training, Integration, and forward-looking phases). Distinct from stage (runtime pipeline) and tier (model vocabulary). 4, separate concern.
• Not a libpostallibpostalAn open-source C address parser used by Pelias. Mailwoman's rule-based v0 and neural classifier supersede it. replacement for everyone. C-language users have libpostallibpostalAn open-source C address parser used by Pelias. Mailwoman's rule-based v0 and neural classifier supersede it.; we're targeting the TypeScript/Node ecosystem.
• Not a research project. We are shipping a library. Novel architectures lose to slightly-tuned standard architectures shipped reliably.
• Not multilingual on day one. The cost of premature internationalization is corpuscorpusThe BIO-labeled training data used to train Mailwoman's neural classifier. Assembled from real sources (OpenAddresses, National Address Database) and synthetic shards (boundary stress, order variants, negative space). Managed by @mailwoman/corpus. chaos and slow trainingtrainingThe process of adjusting a model's parameters so its predictions match labeled examples, by repeatedly measuring error and nudging the weights to reduce it. Distinct from inference, when the trained model is run on new input.. Two localeslocaleThe combination of language and country an address comes from. en-US and fr-FR are the locales Mailwoman ships weights for. is the right scope.