By requesting access, you agree to use the DisStance Similarity Dataset for non-commercial research purposes only, cite the dataset appropriately, and not redistribute the files outside your research group.

DisStance Similarity Dataset

DisStance is a dataset of news-title pairs. Each pair is annotated with a similarity score (0.0 to 1.0, step 0.1) and a stance label (supporting, against, undetermined, not related). It has two parts:

VLDBench pairs: 5,018 pairs built from 838 anchor titles of the VLDBench corpus. Each anchor is paired with up to six candidate titles (two close, two far, two random). The pairs were annotated in three rounds, by six annotators in round 1 and three in rounds 2 and 3.
Supplementary opposing pairs: 169 pairs collected by web scraping to add opposing examples, which are rare in VLDBench. Each pair was annotated by up to six annotators.

Project page: https://ai4debunk.github.io/DisStance-Similarity-Dataset/ · Code: https://github.com/AI4Debunk/DisStance-Similarity-Dataset

Access

The dataset card is public. Downloads are gated: click Access repository, fill in the form, and wait for manual approval. Then log in (hf auth login) before loading the data.

Configs

Config	Rows	Content
`vldbench`	5,018	All VLDBench pairs, one row per pair, with every annotator's label and score
`vldbench_majority`	3,896	Pairs with a strict-majority stance label (the evaluation set)
`vldbench_unanimous`	2,299	Pairs where all annotators gave the same stance and none dismissed the pair
`vldbench_best_pair`	2,718	Pairs on which annotators C and E (the pair with the largest overlap) agree
`vldbench_annotations_long`	20,357	One row per annotator judgment
`supplementary_opposing`	169	All supplementary pairs, with titles
`supplementary_opposing_majority`	141	Supplementary pairs with a strict-majority stance label
`supplementary_opposing_unanimous`	16	Supplementary pairs with full agreement (all opposing)
`supplementary_opposing_best_pair`	102	Supplementary pairs on which annotators B and F agree
`supplementary_opposing_annotations_long`	1,007	One row per annotator judgment

from datasets import load_dataset

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

Pin revision="v2.2" to get exactly the data described here.

Columns (pair-level configs)

Column	Meaning
`main_unique_id`, `db_unique_id`	Anchor and candidate IDs (VLDBench `unique_id` for the VLDBench part)
`db_index`, `db_position`	Candidate slot (0–5) and type: `close`, `far` or `random` (`scraped` for the supplementary part)
`round`, `batch`	Annotation round (1–3)
`related_A` … `related_F`	Stance label of each annotator, or `dismissed`, or empty if the annotator did not see the pair
`similarity_A` … `similarity_F`	Similarity score of each annotator (empty when dismissed or not seen)
`n_annotators`, `n_stance_labels`, `n_dismissed`	Counts over the annotators of the pair
`mean_similarity`	Mean similarity score over the annotators who gave one
`stance_agreement`	`unanimous`, `majority`, `no_majority`, `single_label` or `no_label`
`majority_label`	Stance label chosen by more than half of the stance labels (dismissed excluded), at least two labels required
`in_evaluation`	True when `majority_label` is defined
`main_title`, `comparison_title`	Supplementary part only
`topic`	Supplementary part only: topic family used in the search query (all pairs are `Economy & Finance`)
`main_url`, `comparison_url`	Supplementary part only: address of the anchor and candidate articles
`main_outlet`, `comparison_outlet`	Supplementary part only: website of each article
`main_date`, `comparison_date`	Supplementary part only: publication date (`YYYY-MM-DD`), empty when it could not be found
`main_search_engine`, `comparison_search_engine`	Supplementary part only: news search engine that returned the article
`validation_model`	Supplementary part only: language model that confirmed the pair as opposing before annotation

Agreement configs also have related (the agreed label) and similarity_mean. Each annotator letter refers to the same person in every config.

Labels

Label	Meaning
`supporting`	The two titles support or reinforce the same claim
`against`	The two titles contradict or oppose each other
`undetermined`	Close in meaning, but no clear support or opposition
`not_related`	Not about the same fact, claim or event

dismissed is not a stance. It marks unreadable or out-of-scope pairs and is excluded from agreement and majority labels.

VLDBench headlines

VLDBench configs contain IDs only. To get the headlines, join main_unique_id and db_unique_id with unique_id in VLDBench, for example with extract_annotations.py from the code repository.

Counts

All counts are in CANONICAL_COUNTS_v2.md.

Changelog

v2.2 (2026-11-02): source metadata for the supplementary pairs. The four supplementary pair configs gain topic, main_url, main_outlet, main_date, main_search_engine, comparison_url, comparison_outlet, comparison_date, comparison_search_engine and validation_model (276 of the 338 titles have a publication date, January 2024 to August 2026). Existing columns and rows are unchanged.
v2.1 (2026-10-27): removed duplicate supplementary pairs. The scraping output listed some pairs more than once under different IDs, with the same anchor and candidate titles; 33 copies are removed, keeping the lowest ID number. supplementary_opposing 202 → 169 pairs, supplementary_opposing_majority 166 → 141, supplementary_opposing_unanimous 17 → 16, supplementary_opposing_best_pair 118 → 102. VLDBench configs are unchanged.
v2.0 (2026-10-26): full release of all three annotation rounds.
- Size: VLDBench grows from 665 pairs (111 anchors, round 1 only) to 5,018 pairs (838 anchors, rounds 1–3), built from each annotator's latest annotation file.
- Similarity scores: rounded to the 0.1 grid; 784 scores from one annotator were on a finer grid.
- Duplicates: the same story from two outlets under one anchor (CNN edce44c132 and ABC News 5271935a6e) is now removed under all 10 affected anchors, keeping the lower db_index.
- Annotator letters: the same across all configs. In v1 the supplementary configs used a different assignment, so their letters changed.
- Annotator F: labels on the v1 pairs come from that annotator's latest file and differ from v1 on 93 pairs.
- New configs and columns: *_majority configs, plus the majority_label, in_evaluation and round columns.
- Agreement subset: the VLDBench best annotator pair is now C and E.
v1.1 (2026-10-21): removed one duplicate candidate pair (666 → 665 pairs).
v1.0-666: first release, 666 VLDBench pairs (round 1, 111 anchors).

Downloads last month: 515

Total file size:

824 kB