metadata
dataset_info:
features:
- name: ID
dtype: large_string
- name: statement
dtype: large_string
splits:
- name: train
num_bytes: 1025973
num_examples: 7692
download_size: 413749
dataset_size: 1025973
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
license: odc-by
language:
- en
- hu
- es
- de
- fr
- it
- nl
- pl
pretty_name: probvaa
size_categories:
- 1K<n<10K
ProbVAA
ProbVAA is a compilation of voting advice applications (VAAs).
Coding information for the IDs
String with the following format:
CountryCode + _ + StatementID + _ + Original|Question|ParaphraseID|Opposite|Negation|Country-Agnostic|Translation + _ + LanguageCode
e.g. ch_15_0000001_fr
- Original, opposite, question, negation, country-agnostic and translation are binary.
- StatementID is the ID of the statement in the original dataset.
- ParaphraseID = 0 means it's not a paraphrase. From 1 on it's the ID.
- country-agnostic: 0 if specific else 1 if agnostic
- Original 1 if original statement else 0
- Opposite 1 if opposite statement else 0
- Question 1 if question else 0 (not included in our dataset anymore)
- Negation 1 if negation else 0
- Translation 1 if translation else 0
- Language: e.g. en, es, de...
- CountryCode: e.g. ch, de, fr, it
Full information about the creation of the dataset in this paper:
@article{ceron-etal-2024-beyond,
title = "Beyond Prompt Brittleness: Evaluating the Reliability and Consistency of Political Worldviews in {LLM}s",
author = "Ceron, Tanise and
Falk, Neele and
Bari{\'c}, Ana and
Nikolaev, Dmitry and
Pad{\'o}, Sebastian",
journal = "Transactions of the Association for Computational Linguistics",
volume = "12",
year = "2024",
address = "Cambridge, MA",
publisher = "MIT Press",
url = "https://aclanthology.org/2024.tacl-1.76/",
doi = "10.1162/tacl_a_00710",
pages = "1378--1400",
abstract = "Due to the widespread use of large language models (LLMs), we need to understand whether they embed a specific ``worldview'' and what these views reflect. Recent studies report that, prompted with political questionnaires, LLMs show left-liberal leanings (Feng et al., 2023; Motoki et al., 2024). However, it is as yet unclear whether these leanings are reliable (robust to prompt variations) and whether the leaning is consistent across policies and political leaning. We propose a series of tests which assess the reliability and consistency of LLMs' stances on political statements based on a dataset of voting-advice questionnaires collected from seven EU countries and annotated for policy issues. We study LLMs ranging in size from 7B to 70B parameters and find that their reliability increases with parameter count. Larger models show overall stronger alignment with left-leaning parties but differ among policy programs: They show a (left-wing) positive stance towards environment protection, social welfare state, and liberal society but also (right-wing) law and order, with no consistent preferences in the areas of foreign policy and migration."
}