LLM-generated parliamentary text detection

This data set consists of human-written parliamentary texts and LLM-generated version of such texts in two languages, English and Swedish. The human-written texts have been obtained from the UK parliament (Hansard) [1] and Sweden's parliament (Riksdag) [2]. The LLM-generated texts were generated by Gemini 3.1 Flash-Lite and GPT-5 mini. The data is published as a supplement to the paper Detecting undisclosed LLM-generated content in parliamentary texts [3].

The texts have been tokenized (as described in the paper) such that each sample can be split on whitespace characters to obtain a tokenized text. 

The base data sets consist of Training, Validation, and Test sets with paragraphs of at least 30 tokens.These sets are suitable for training and evaluating a text classifier that distinguishes between human-written (label 0) and LLM-generated (label 1) texts. The files are structured such that every line contains a class label (0 or 1) and a text, separated by a tab character.

The forward test sets consist mainly of texts that were published after the public release of LLMs in 2022. The year 2021 is also included in the forward set as a subset to test for false positive detections. For this set, the text samples do not have ground truth labels.

[1] Source: Hansard, https://hansard.parliament.uk/. Contains Parliamentary information licensed under the Open Parliament Licence v3.0. License: https://www.parliament.uk/site-information/copyright/open-parliament-licence/

[2] Source: The Swedish Parliament, see https://www.riksdagen.se/sv/dokument-och-lagar/riksdagens-oppna-data/dokument/ and https://www.riksdagen.se/sv/dokument-och-lagar/riksdagens-oppna-data/anvandarstod/anvandningsvillkor/

[3] Suvanto, M., McGlinchey, A., Wahde, M., & Barclay, P. J. (2026). Detecting undisclosed LLM-generated content in parliamentary texts. (Manuscript submitted for publication)