Million English Numbers
A list of a million American English numbers, under a AGPL 3.0 license. This datasheet is inspired by Datasheets for Datasets.
Sample
$ tail -n 5 million-english-numbers
nine hundred ninety nine thousand nine hundred ninety five
nine hundred ninety nine thousand nine hundred ninety six
nine hundred ninety nine thousand nine hundred ninety seven
nine hundred ninety nine thousand nine hundred ninety eight
nine hundred ninety nine thousand nine hundred ninety nine
Motivation
This dataset was created as a toy sample of text for use in natural language processing, in machine learning.
The goal was to create small samples of text with minimal variation and results that could be easily audited (observe how often the model predicts "eighty twenty hundred three ten forty").
This is original research, produced by the linguistic model in the NodeJS package written-number by Pedro Tacla Yamada, freely available on npm.
The estimated cost of creating the dataset is minimal, and subsidized with private funds.
Composition
The instances that comprise the dataset are spelled-out integers, in colloquial Mid-Atlantic American English, identifiable to a speaker born around the year 2000.
There are one million instances, from 0 to 999999 consecutively.
The instances consist of ASCII text, delimited by line feeds.
Counting lines from zero, the line number of each instance is its integer value.
No information is missing from each instance.
In the related fast.ai HUMAN_NUMBERS dataset, the split is between 1-7999, and 8001-9999. A user may elect to split this dataset similarly, with the last percentages of lines used for validation or testing.
There are no known errors or sources of noise or redundancies in the dataset.
The dataset is self-contained.
The dataset is not confidential, and its method of generation is public as well.
The dataset will probably not be offensive / insulting / threatening / anxiety-inducing to many people. The numerologically-minded may wish to exercise discernment when choosing which numbers to use: all of the auspicious numbers, all of the inauspicious numbers, all of the meaningful numbers, for all numerological traditions, are included in this dataset, without any emphasis or warnings besides sequential ordering.
The dataset does not relate to people, except by using human language to express integers.
Collection
The data was directly observed from the written-number npm package.
To rebuild this dataset, run docker run -e MAXINT=1000000 -e WN=written-number -w /x node sh -c 'npm i $WN 2>1 >/dev/null; node -e "const w=require(process.env.WN);for(i=0;i<process.env.MAXINT;i++) console.log(w(i,{noAnd: true}))" | tr "-" " "' on any x86 machine with Docker. Manual spot-checking confirmed the results.
This is a subset of the set of integers, in increasing order, with no omissions, starting from zero.
This was collected by one individual, writing minimal code, using free time donated to the project.
The data was collected at one point in time, using colloquial Mid-Atlantic American English. The idea of integers including zero is long-standing, and dates back to Babylonians in 700 BCE, the Olmec and Maya in 100 BCE, Brahmagupta in 628 CE.
There was no IRB involved in the making of this data product.
The instances individually do not relate to people.
Preprocessing
The default output of the version of written-number puts a hyphen between the tens and ones place, and this hyphen was translated into a space in the output.
Further, the default conjunction and between the hundreds and tens place was removed, as visible above in the sample (nine hundred ninety).
This raw data was not saved.
The code to regenerate the raw data, and the code to run the preprocessing, is available in this datasheet.
Uses
This dataset has not been used for any tasks already. It has been inspired by a smaller fast.ai dataset used pedagogically for training NLP models.
The dataset could also be used in place of the original code that generated it, if someone desired a list of human-readable numbers in this dialect of English. The dataset could also be used as a normative spelling of integers (to correct someone writing "fourty" for instance). The dataset could also be used, as an artifact of language, could be used to establish normative language for reading integers.
The dataset is composed of only one of the many languages and dialects that written-number produces.
A native user of another dialect might elect to change language or dialect, for easier auditing of the output of the language model trained on the numbers.
Specifically, someone might expect to see nine lakh ninety nine thousand nine hundred ninety nine instead of nine hundred ninety nine thousand nine hundred ninety nine as the last line of the sample above.
It is important to not use this dataset as a normative spelling of integers, especially to impose American English readings of integers on speakers of other dialects of English.
Distribution
This dataset is distributed worldwide.
It is available on Huggingface, at https://huggingface.co/datasets/lsb/million-english-numbers .
It is currently available.
The license is AGPL 3.0.
The library written-number is available under the MIT license, and its output is not currently restricted by license.
No third parties have imposed any restrictions on the data associated with these instance of written numbers.
No export controls or other regulatory restrictions currently apply to the dataset or to individual instances in the dataset.
Maintenance
Huggingface is currently hosting the dataset, and @lsb is maintaining the dataset.
Contact is available via pull-request, and via email at [email protected] .
There are currently no errata, and the full edit history of the dataset is available in the git repository in which this datasheet is included.
This dataset is not expected to frequently update. Any users of the dataset may elect to git pull any updates.
The data does not relate to people, and there are no limits on the retention of the data associated with the instances.
Older versions of the dataset continue to be supported and hosted and maintained, through the git repository that includes the full edit history of the dataset.
If others wish to extend or augment or build on or contribute to the dataset, a mechanism available is to upload additional datasets to Huggingface.