README.md

metadata

license: apache-2.0
task_categories:
  - text-generation
language:
  - en
tags:
  - generics
pretty_name: MGen
size_categories:
  - 1M<n<10M

MGen Dataset: Millions of Generics in Context

This repository hosts the MGen dataset: a collection of millions of naturally occuring generics and quantified sentences in context.

This dataset is designed to be a resource for the empirical study of generic sentences. Find the details of the construction in the SCiL 2025 paper or the blogpost.

Data Structure

The dataset is released in several .csv files.

• sentences.csv This is the main file, containing only the generic and quantified sentences with some metadata, as described below.

Please note that to properly open this csv with pandas you need to specify the lineterminator: mgen = pd.read_csv('data/mgen/sentences.csv', lineterminator='\n'). Without this, the data will not open up properly!

Column Name	Description
id	Unique id of the sentence, doc_id + the cardinality of the sentence within that doc.
doc_id	Unique document id (hash of the raw document text, more details below)
text	Target sentence (string).
quantifier	Qunatifier that appears in the main clause of the sentence, one of: gen, all, most, many, some, few, no, often, generally, typically, usually, normally
source	Source of the original document. One of arxiv, peS2o, slimpajama, refinedweb, pile-uncopyrighted.
score	Score of the RoBERTa-based classifier for genericity. Sentences are selected to have greater score than 0.8.
n_words	Number of words in the sentence (counted by splitting whitespaces).
quantifier_position	Quantifier position. Either first_word or middle if the quantifier is after the bare plural (i.e. tigers are normally striped).
quantifier_category	One of gen, determiner, adverbial.

• Documents with the contexts of the generic/quantified sentences from each of the five sources.

Column Name	Description
doc_id	Unique document id (hash of the raw document text, more details below)
text	Context text (string).
source	Source of the document.

The doc_id is calculated by hashing the raw string of the whole documents from ZYDA.

The following python snippet calculates the hash (doc_id) from the text of a document: hashlib.sha1(text).hexdigest().

Note that this repository does not currently contain the context documents, to download those please see the blogpost.

Recommended Usage

It is recommended to sample first from sentences.csv, finding those sentences of interest (because of topic, quantifier or other reasons), and then find the contexts of this subset. Because the contexts are long, often corresponding to websites or research papers, the documents .csv are very heavy and slow to work with.

Cite

If you use this dataset, please please please cite the paper! Also feel free to contact me for any doubts or additional data!

@article{cilleruelo2025mgen,
  author    = {Cilleruelo, G. and Allaway, E. and Haddow, B. and Birch, A.},
  title     = {MGEN: Millions of Naturally Occurring Generics in Context},
  journal   = {Society for Computation in Linguistics},
  volume    = {8},
  number    = {1},
  pages     = {10},
  year      = {2025},
  doi       = {10.7275/scil.3147}
}

This work was funded by the European Union’s Horizon Europe (HE) Research and Innovation programme under Grant Agreement No 101070631 and from the UK Research and Innovation (UKRI) under the UK government’s HE funding grant No 10039436.