Dataset · release v2.2

DisStance

News-title pairs annotated with a similarity score (0.0 to 1.0, step 0.1) and a stance label. One row is one pair, with every annotator's judgment.

Dataset on Hugging Face Code on GitHub

VLDBench pairs (IDs only) and supplementary opposing pairs (titles included), with their evaluation and agreement subsets.

IDs only

VLDBench pairs

5,018 pairs from 838 anchor titles, annotated in three rounds · titles rebuilt from VLDBench

Titles included

Supplementary opposing

169 scraped pairs that add opposing examples · usable as is

Pick the right set

Goal	Config	Rows
All VLDBench pairs, every annotator	`vldbench`	5,018
Evaluation set (strict-majority stance label)	`vldbench_majority`	3,896
All annotators agree	`vldbench_unanimous`	2,299
Best annotator pair agrees (C and E)	`vldbench_best_pair`	2,718
All supplementary pairs, with titles	`supplementary_opposing`	169
Supplementary evaluation set	`supplementary_opposing_majority`	141
Supplementary, full agreement	`supplementary_opposing_unanimous`	16
Supplementary, best pair agrees (B and F)	`supplementary_opposing_best_pair`	102
One row per judgment	`*_annotations_long`	20,357 / 1,007

The reference label of a pair is majority_label: the stance chosen by more than half of its annotators, dismissed judgments excluded. Agreement and majority configs also give related and similarity_mean. The long tables are meant for agreement studies, not for training.

Labels

supporting against undetermined not_related dismissed = skipped, not a stance

Access

The dataset card on Hugging Face is public. Downloads are gated: click Access repository on the dataset page, fill in the form and wait for manual approval, then log in with hf auth login.

The anchor and candidate titles of the VLDBench part come from VLDBench, which you download separately.

Load

from datasets import load_dataset

REPO = "AI4Debunk/DisStance-Similarity-Dataset"
pairs = load_dataset(REPO, "vldbench", split="train", revision="v2.2")
evaluation = load_dataset(REPO, "vldbench_majority", split="train", revision="v2.2")
opposing = load_dataset(REPO, "supplementary_opposing_majority", split="train", revision="v2.2")

The VLDBench configs contain IDs only. extract_annotations.py in the code repository adds the VLDBench headlines and text:

python extract_annotations.py --config vldbench_majority --format csv --output vldbench_majority.csv

Earlier releases remain available as revision="v1.0-666" and revision="v1.1".

How to cite

Placeholder (update when the paper is public):

Mokhtari, M. E. A., et al. (2026). DisStance: A Dataset for Title-Level Similarity and Stance Relations in Disinformation Detection. Preprint forthcoming.

@misc{disstance2026,
  title        = {DisStance: A Dataset for Title-Level Similarity and Stance Relations in Disinformation Detection},
  author       = {Mokhtari, Mohammed El Amine and others},
  year         = {2026},
  howpublished = {Dataset},
  note         = {Citation placeholder, to be replaced with the paper entry},
  url          = {https://huggingface.co/datasets/AI4Debunk/DisStance-Similarity-Dataset}
}