OpenStreetMap attribution and data licensing

Mailwoman uses data from sources with different licenses and provenance. OpenStreetMap is one of them. We value the work of 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. contributors and aim to make it clear when OSMOpenStreetMap (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. data helped train 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.', when OSMOpenStreetMap (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. is being used to look up an address, and which obligations apply to an OSMOpenStreetMap (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.-derived database.

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. attribution © 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. contributors. 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. data is available under the Open Database License (ODbL) 1.0. See openstreetmap.org/copyright.

Two different uses of OpenStreetMap

Mailwoman makes a distinction between its address-understanding modelsneural 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.' and optional address-retrieval datasets.

1. Address-understanding models

Some released Mailwoman modelsneural 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.' have learned from address examples derived from 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.. That 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. helps them recognize address patterns across countries and scripts.

A trained 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 not the same artifact as the 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. address dataset used to train it. Our interpretation, informed by the ODbL and OpenStreetMap Foundation's machine-learning attribution guidance, is that model weightsmodel weightsThe learned parameters of the neural classifier, shipped as ONNX files in the @mailwoman/neural-weights-* packages. Weights are locale-specific bundles that include the model, tokenizer, and a model-card.json metadata file. can have their own license while their documentation credits OSMOpenStreetMap (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. and the underlying OSMOpenStreetMap (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.-derived 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. database remains subject to its own ODbL obligations.

For modelsneural 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 have OSMOpenStreetMap (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. in their effective 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. ancestry, we include attribution in 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 rights and provenance documentation. We also maintain the records needed to identify the relevant 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. inputs and satisfy applicable access/reconstruction requirements for derivative databases.

Not enabling the optional OSMOpenStreetMap (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. retrieval 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. does not erase 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.''s 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. history. 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.' can require OSMOpenStreetMap (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. attribution even when your application never installs an OSMOpenStreetMap (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. address-point database.

Ordinary predictions or parsed address components are not automatically licensed under ODbL because 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.' was trained on OSMOpenStreetMap (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.-derived data. Systematically extracting or reproducing substantial portions of OSMOpenStreetMap (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. data is a different use case.

2. Optional address retrieval

Mailwoman can use an additional address-point 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. built from an 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. extractextractA partial output file of the corpus build, written in Parquet format. The training pipeline streams extracts row by row.. This 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. improves retrieval and matching in places where other datasets are incomplete.

This is opt-in. The OSMOpenStreetMap (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. adapter is software. The OSMOpenStreetMap (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. data it processes is separately licensed. An OSMOpenStreetMap (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.-backed 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. leaves the license of Mailwoman's application code unchanged, and it does not automatically relicense your independently sourced address records.

The intended workflow is to download an OSMOpenStreetMap (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. extractextractA partial output file of the corpus build, written in Parquet format. The training pipeline streams extracts row by row. from a provider such as Geofabrik, choose a 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., and build a local address-point SQLite 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.. The resulting OSMOpenStreetMap (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.-derived database is identified as ODbL-1.0 and kept distinguishable from other layerslayerOne 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.. Check your installed Mailwoman version for supported build commands and source availability.

What changes when you enable OSM retrieval?

Situation	What to expect
You use Mailwoman without an OSMOpenStreetMap (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. retrieval 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.	Mailwoman by itself fetches and queries no OSMOpenStreetMap (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. address-point 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.. 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.' may still require OSMOpenStreetMap (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. 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. attribution.
You build and use an OSMOpenStreetMap (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.-derived 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. internally	Keep its ODbL license, attribution and provenance. Internal use generally does not by itself trigger the ODbL public-use share-alike requirements.
You expose an OSMOpenStreetMap (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.-backed geocoder or application publicly	Credit 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. in a place users can reasonably find. Additional database-access duties may apply if you use a modified OSMOpenStreetMap (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.-derived database.
You save an ordinary individual address match alongside a customer record	This does not automatically make the customer database ODbL. The OSMF's 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. guidance distinguishes ordinary insubstantial results from derivative databases.
You systematically incorporate, alter or reconstruct substantial OSMOpenStreetMap (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. data in another database	The resulting database may be a Derivative Database subject to ODbL share-alike and database-access obligations.

These distinctions depend on what your system does with the data. How you labelcomponent tagOne of the 25 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. or store databases decides none of them.

Your data and OSM data remain distinguishable

Our retrieval design keeps source provenance attached to the datasets and results that use them. An OSMOpenStreetMap (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.-derived SQLite 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. is treated as OSMOpenStreetMap (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.-derived data. Another independently sourced 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. retains its own license.

The OSMF Geocoding Guideline expressly discusses ordinary 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. results stored with proprietary records. It states that those results generally do not trigger share-alike for the surrounding database when they are insubstantial and are not aggregated to reproduce a substantial portion of OSMOpenStreetMap (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..

This is not a blanket exemption for every enrichment workflow. If you use OSMOpenStreetMap (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. to populate, correct or systematically assemble a larger database, the nature and scale of those changes matter. Independent data and OSMOpenStreetMap (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.-derived data remain distinguishable when each has clear provenance.

Attribution when you operate a public service

If your service exposes OSMOpenStreetMap (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.-backed lookups, include a visible or conventionally located credit such as:

Address data © OpenStreetMap contributors, available under ODbL 1.0.

This may appear in the application, API documentation, service credits or another location users would ordinarily expect. Preserve provenance in your application where practical. A source field inside an individual API response should not be your only attribution mechanism for a public OSMOpenStreetMap (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.-backed service.

Each Mailwoman HTTP surface keeps a result's openstreetmap:<cc> source stamp when an OSMOpenStreetMap (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. 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. you attached supplied its coordinate, and the Nominatim and Photon surfaces add the credit "© 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. contributors, ODbL 1.0, https://www.openstreetmap.org/copyright" to that result. A result from any other source never has the stamp or the credit, and a deployment without an OSMOpenStreetMap (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. 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. serves no OSMOpenStreetMap (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. data. HTTP APIs lists the field each surface uses.

Distribution of an OSM-derived database

The ODbL permits commercial use. It also contains conditions for publicly using or distributing modified OSMOpenStreetMap (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. databases.

In particular, §4.6 can require an offer to provide recipients with either the relevant derivative database in machine-readable form, or the alteration file/method that recreates it, including applicable additions. The details depend on whether you publicly use a Derivative Database or a Produced Work made from one.

Mailwoman's licensing approach is to keep these obligations attached to the relevant OSMOpenStreetMap (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.-derived database rather than claim ownership over OSMOpenStreetMap (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. data or imply that every other dataset and application component must use the same license.

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. data retains its license requirements when you use it. Mailwoman's source attribution does not replace the obligations applicable to a service you operate.

How to inspect your sources

For a generated OSMOpenStreetMap (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. 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., look for its data 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. manifest. It records the license, the credit, the Geofabrik URL and SHA-256 of the extractextractA partial output file of the corpus build, written in Parquet format. The training pipeline streams extracts row by row. it was built from, and the code revision that built it. mailwoman doctor lists each address-point database under $MAILWOMAN_DATA_ROOT/db/osm with the license and credit its manifest records, and mailwoman data bom reports each one as its own component. To decline share-alike data, set MAILWOMAN_REFUSE_OBLIGATIONS=share-alike: mailwoman data pull then refuses share-alike bundlesevidence bundleThe pair of retrieval-augmented input channels (street-type + locality-surface) that feed lexicon membership as soft per-token evidence alongside the text. Shipped in 6.7.0; trained natively from step 0 in the from-scratch base line., and the geocode session refuses to start while an 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. database is on disk, naming the database and its recorded provenance.

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.' rights files and model cardsmodel cardA JSON metadata file (model-card.json) shipped with each weights bundle. It declares the model version, lineage, label set, required inference channels (anchor, gazetteer), calibration data, and training provenance. are separate from the retrieval-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. manifest. They report 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. lineage and notices associated with that 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 weightsparameterA single learned number inside a model — one weight or bias. Mailwoman's encoder has roughly 30 million of them; training is the search for good values..

Which released models trained on OpenStreetMap

The @mailwoman/neural-weights-en-us 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.' version 10.1.0 (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.' v7.2.0-address-systems) trained on corpuscorpusThe BIO-labeled training data used to train Mailwoman's neural classifier. Assembled from real sources (OpenAddresses, National Address Database) and synthetic extracts (boundary stress, order variants, negative space). Managed by @mailwoman/corpus. v0.7.2-address-systems, whose osm-ru, osm-in, osm-id, osm-ng, osm-pk, osm-vn and osm-bd sources derive from the 2026-10-02 Geofabrik extractsextractA partial output file of the corpus build, written in Parquet format. The training pipeline streams extracts row by row.. One audited 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. epoch emitted 103,322 of its 1,000,000 rows from those sources and 15,609 from two registersinput modeThe Decision-A register switch: 'fragmented' (human-typed fragments — feeds the evidence channels) vs 'formatted' (complete records — runs the trained absence identity). Explicit on CLI/API; per-endpoint defaults (batch→formatted, autocomplete→fragmented); kind-derived otherwise. rendered from OSMOpenStreetMap (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. tuples. Every localelocaleThe combination of language and country an address comes from. en-US and fr-FR are the locales Mailwoman ships weights for. overlay that decodes through that graph (en-gb, en-au, en-nz, en-in, de-de, es-es, it-it, fr-fr) inherits this lineage.

The npm packages published as 10.1.0, 10.2.0, 10.2.1, 10.2.2, 10.3.0 and 10.4.0 ship these weightsparameterA single learned number inside a model — one weight or bias. Mailwoman's encoder has roughly 30 million of them; training is the search for good values. without an 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. notice in their LICENSE.md, PROVENANCE.json, README.md or model-card.json. A published npm tarball cannot be altered, so those versions remain as published and this page is the notice for them. Release 10.5.0, published 2026-10-17, is the first whose rights files derive the attribution from 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. record: each weightsparameterA single learned number inside a model — one weight or bias. Mailwoman's encoder has roughly 30 million of them; training is the search for good values. package's LICENSE.md prints the 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. statement over the sources that reached the trainer, PROVENANCE.json records the effective 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. sources with their row counts, and the en-us README and model cardmodel cardA JSON metadata file (model-card.json) shipped with each weights bundle. It declares the model version, lineage, label set, required inference channels (anchor, gazetteer), calibration data, and training provenance. contain the notice in prose. The frozen 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. record and the effective 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. manifest are committed in the repository under packages/corpus/data/training-manifests/.

The training extract

The 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.-derived 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. extractextractA partial output file of the corpus build, written in Parquet format. The training pipeline streams extracts row by row. of corpuscorpusThe BIO-labeled training data used to train Mailwoman's neural classifier. Assembled from real sources (OpenAddresses, National Address Database) and synthetic extracts (boundary stress, order variants, negative space). Managed by @mailwoman/corpus. v0.7.2-address-systems is published under ODbL 1.0 at huggingface.co/datasets/sister-software/mailwoman-osm-training-extract. It is the offer that ODbL section 4.6(a) requires for the weightsparameterA single learned number inside a model — one weight or bias. Mailwoman's encoder has roughly 30 million of them; training is the search for good values. above: the derivative database in machine-readable form, as parquetParquetThe open columnar file format the corpus is written and streamed in. The training pipeline reads extracts row-by-row from Parquet. files of the rows the trainer read. It holds 2,619,226 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. rows in 32 files. Twenty-nine of them hold the 2,572,934 rows of the eight osm-* sources, the earlier osm source and the two registersinput modeThe Decision-A register switch: 'fragmented' (human-typed fragments — feeds the evidence channels) vs 'formatted' (complete records — runs the trained absence identity). Explicit on CLI/API; per-endpoint defaults (batch→formatted, autocomplete→fragmented); kind-derived otherwise. rendered from OSMOpenStreetMap (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. tuples, each with the row count and sha256 digest the corpuscorpusThe BIO-labeled training data used to train Mailwoman's neural classifier. Assembled from real sources (OpenAddresses, National Address Database) and synthetic extracts (boundary stress, order variants, negative space). Managed by @mailwoman/corpus. manifest lists. The other three hold spliced-sub-venue rows, described below. The dataset's MANIFEST.json lists each file with its row count and sha256 digest, and its README states the Geofabrik extractextractA partial output file of the corpus build, written in Parquet format. The training pipeline streams extracts row by row. dates and the repository revision that produced the files. The en-us package's PROVENANCE.json links the dataset under data_access_offers.

The spliced-sub-venue recipe splices 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. and sub-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. names onto address skeletons. Its rows record no name source, so each row was matched against the name lists its localelocaleThe combination of language and country an address comes from. en-US and fr-FR are the locales Mailwoman ships weights for. drew from. The en-GB, de-DE, es-ES and ca-ES rows drew only from 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. extractsextractA partial output file of the corpus build, written in Parquet format. The training pipeline streams extracts row by row., en-US only from Overture Places, and fr-FR from both. Of the corpuscorpusThe BIO-labeled training data used to train Mailwoman's neural classifier. Assembled from real sources (OpenAddresses, National Address Database) and synthetic extracts (boundary stress, order variants, negative space). Managed by @mailwoman/corpus.'s 120,000 rows, 73,039 have an 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. name: 71,598 an 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. 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, 547 an 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. sub-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. string beside an Overture 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, and 894 a name both lists hold. The other 32,658 rows have an Overture Places name and 14,303 have no pooled name.

The dataset publishes 46,292 of those 73,039 rows, every one outside the en-GB localelocaleThe combination of language and country an address comes from. en-US and fr-FR are the locales Mailwoman ships weights for.. The 26,747 en-GB rows were withdrawn on 2026-10-17. Their address skeletons (house numberhouse numberThe numeric or alphanumeric identifier of a building on a street. Mailwoman's house_number component; its position relative to the street name flips between locales., streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels., localitylocalityThe city / town / settlement component of an address: a populated place sitting between region and neighborhood in the hierarchy., regionregionThe first-level administrative subdivision of a country — a US state, a French region, a province. The component between country and locality. and postcodepostcodeThe country-specific postcode (US ZIP, French code postal, etc.). Mailwoman handles postcode parsing entirely by rule classifier — a regex problem, not an ML one.) were fields of HM Land Registry Price Paid Data. Its download page excludes those Address Data fields from the Open Government Licence. Royal Mail and Ordnance Survey permit them only for personal or non-commercial use and for displaying residential property price information. The earlier revisions of the dataset that held those rows were removed from its history. The Great Britain extractextractA partial output file of the corpus build, written in Parquet format. The training pipeline streams extracts row by row. file stays published. It holds the whole 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. name list that localelocaleThe combination of language and country an address comes from. en-US and fr-FR are the locales Mailwoman ships weights for. drew from, including every name the en-GB rows had.

The published rows' skeletons come from the Base Adresse Nationale (fr-FR, Licence Ouverte 2.0), the Berlin and Sachsen address registersinput modeThe Decision-A register switch: 'fragmented' (human-typed fragments — feeds the evidence channels) vs 'formatted' (complete records — runs the trained absence identity). Explicit on CLI/API; per-endpoint defaults (batch→formatted, autocomplete→fragmented); kind-derived otherwise. through OpenAddressesOpenAddresses (OA). A global open aggregation of address points collected from many official sources. A primary source of component-supervised training data outside proprietary registries. (de-DE, Datenlizenz Deutschland Zero 2.0 and Namensnennung 2.0), and CartoCiudad through OpenAddressesOpenAddresses (OA). A global open aggregation of address points collected from many official sources. A primary source of component-supervised training data outside proprietary registries. (es-ES and ca-ES, CC BY 4.0). The dataset also holds the four 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. extractsextractA partial output file of the corpus build, written in Parquet format. The training pipeline streams extracts row by row. the recipe drew names from, and its README states the method, the counts and the credit each skeleton source requires. The recipe now records each row's name sources in its output, and the effective 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. manifest lists spliced-sub-venue among the sources that owe the 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. attribution.

References and our interpretation

We base our approach on the following publicly available documents:

• Open Database License 1.0 — full text, especially §§2.3, 3.1 and 4.3–4.6.
• OpenStreetMap Foundation Attribution Guideline, including its machine-learning section.
• OpenStreetMap Foundation Geocoding Guideline.
• OpenStreetMap Foundation Collective Database Guideline.
• OpenStreetMap Foundation FAQ on the license and legal questions.

Our position: model weightsmodel weightsThe learned parameters of the neural classifier, shipped as ONNX files in the @mailwoman/neural-weights-* packages. Weights are locale-specific bundles that include the model, tokenizer, and a model-card.json metadata file., adapter software, independently sourced customer records and OSMOpenStreetMap (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.-derived address databases are four different artifacts, each under its own terms. We attribute 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. when its data contributes, preserve source provenance and respect the ODbL conditions on OSMOpenStreetMap (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.-derived databases, including applicable availability obligations.

This page explains Mailwoman's licensing interpretation and intended behavior. Legal advice for your own situation comes from your own adviser, and rulings on the license come from the 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. Foundation. The ODbL text governs your use of OSMOpenStreetMap (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. data.