Dataset Card for worldwide-addresses

Dataset description

Dataset Summary

This dataset is a collection of annotated international addresses containing over 750,000,000 addresses from 240 countries in over 100 languages. It has been created from the data gathered and provided by libpostal, an international street address parsing package. The original purpose of this dataset was to develop a state-of-the-art neural network-based international address parser named deepparse.

The dataset is structured in a way that each country has its own configuration. Therefore the data can be loaded by specifying a country's ISO 3166-1 alpha-2 code as a config name:

from datasets import load_dataset

ds = load_dataset("deepparse/worldwide-addresses", "us")

print(ds["train"][0])

Supported Tasks and Leaderboards

• token-classification: The dataset can be used to train models for token classification, which consists of assigning a class to each token in a text sequence. In this case, the dataset can be used to train an address parser that is able to identify the different elements of an address such as a street name or a postal code.

Languages

Each country's addresses can be expressed in multiple languages. For example, the us data contains addresses from the United States which can be in a language other than english (e.g: in spanish). Since this is a Parquet-based dataset and the language is included in each sample's data fields, you can specify a filter to exclusively load addresses for a specific language. This is done by specifying the language's ISO 639-3 code like this:

from datasets import load_dataset

lang_iso = "eng"

# Only load addresses in English (eng)
lang_filter = [("Language", "==", lang_iso)]

ds = load_dataset("deepparse/worldwide-addresses", "us", filters=lang_filter)

Dataset Structure

Data Instances

Each sample is formatted in the following way:

{
  'Address': 'Douglas County Minnesota 56332',
  'Tags': ['County', 'County', 'Province', 'PostalCode'],
  'Language': 'eng'
}

Data Fields

The dataset contains three fields:

• Address: this is a String representing the address itself. There's no punctuation, so each word in the address is seperated by a whitespace. When training a model for token-classification, this would constitute the input.
• Tags: these are the annotations that associate a standardized address element for each word in the address. It is a list with the same length as the number of whitespace seperated words in the Address field. The tags are defined as follows:
  • StreetNumber: a house or a building number.
  • StreetName: the name of the street.
  • Unit: an apartment or a unit number.
  • Suburb: an unofficial neighbourhood name.
  • District: the name of a neighbourhood which has official administrative boundaries.
  • PostalCode: standard postal code which vary per country.
  • Municipality: the name of a city.
  • Province: the name of a sub-national division within a country.
  • County: the name of a major administrative area within a province.
  • Country: the name of a country.
• Language: the ISO 639-3 code representing the language in which the address is written.

Dataset Creation

This dataset was curated and adapted from an international addresses dataset published by libpostal. For more details on the dataset creation process visit their post.

Additional information

Licensing Information

This dataset is shared under a CC-BY-4.0 license.

Citation Information

If you use this dataset please cite the following:

@misc{worldwide-addresses,
    author = {Marouane Yassine and David Beauchemin},
    title  = {{Structured Multinational Address Data}},
    year   = {2026},
    note   = {\url{https://huggingface.co/datasets/deepparse/worldwide-addresses}}
}

Contributions

Thanks to @albarrentine for creating and sharing the original dataset used to train libpostal's models, and for developping the package itself. It is quite an impressive piece of work!