You need to agree to share your Hugging Face information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
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.
Log in or Sign Up to review the conditions and access this dataset content.
DisStance Similarity Dataset
Hugging Face access
The Hub repository amine0110/DisStance-Similarity-Dataset is public: anyone can read the dataset card, configs, and documentation. Downloads are gated — open the dataset page, click Access repository, complete the form, and wait for the maintainer to approve your request manually (access is not auto-granted). After approval, authenticate with your Hugging Face account (for example huggingface-cli login) before running the load_dataset examples below.
Dataset website
When this repository is public, the project page is served from docs/ via GitHub Pages:
https://amine0110.github.io/DisStance-Similarity-Dataset/
While the repository remains private, that Pages site is only visible to collaborators with access.
Human similarity and relationship annotations for news headline pairs. DisStance is a multi-source release. Packaging rule (2026-09-17): one row = one unique pair. Do not treat multi-annotator long tables as i.i.d. samples.
Which Hugging Face config should I load?
| Goal | Config | Rows |
|---|---|---|
| All unique VLDBench pairs (per-annotator columns) | vldbench (default) |
666 |
| VLDBench pairs where all labeling annotators agree | vldbench_unanimous |
309 |
| VLDBench pairs where the best annotator pair agrees (C + F) | vldbench_best_pair |
510 |
| All unique scrape pairs (titles included) | supplementary_opposing |
202 |
| Scrape full agreement | supplementary_opposing_unanimous |
17 |
| Scrape best annotator pair (B + C) | supplementary_opposing_best_pair |
118 |
| IAA / raw multi-annotator rows only | *_annotations_long |
archived |
On agreement configs, related is the shared stance label and similarity_mean is the mean similarity across the agreeing annotators.
from datasets import load_dataset
# VLDBench unique pairs + agreement subsets
pairs = load_dataset("amine0110/DisStance-Similarity-Dataset", "vldbench", split="train")
uni = load_dataset("amine0110/DisStance-Similarity-Dataset", "vldbench_unanimous", split="train")
best = load_dataset("amine0110/DisStance-Similarity-Dataset", "vldbench_best_pair", split="train")
# Supplementary opposing unique pairs + agreement subsets
opp = load_dataset("amine0110/DisStance-Similarity-Dataset", "supplementary_opposing", split="train")
opp_uni = load_dataset("amine0110/DisStance-Similarity-Dataset", "supplementary_opposing_unanimous", split="train")
opp_best = load_dataset("amine0110/DisStance-Similarity-Dataset", "supplementary_opposing_best_pair", split="train")
Sources and redistribution
| Source | Config family | Headline text in release? |
|---|---|---|
| VLDBench | vldbench* |
No (IDs only; reconstruct locally with VLDBench + extract_annotations.py) |
| Supplementary opposing scrape | supplementary_opposing* |
Yes (main_title, comparison_title) |
Labels
| Label | Meaning |
|---|---|
supporting |
Supports or reinforces the main claim |
against |
Contradicts or opposes the main claim |
undetermined |
Same topic/event, no clear support or opposition |
not_related |
Clearly unrelated |
dismissed is a reject/skip decision, not a fifth stance. Pairs with any dismissed vote are excluded from unanimous sets.
Data files in this repository
See UNIQUE_PAIR_RELEASE.md and data/*_pairs.parquet, data/*_unanimous.parquet, data/*_best_pair.parquet. Long form: data/*_annotations_long.parquet (IAA only).
Reconstruct VLDBench text locally
Scrape configs already include titles. For VLDBench configs, join IDs to a local VLDBench download with extract_annotations.py. Pick the subset with --config:
pip install -r requirements.txt
# Unique pairs (related_A..F, similarity_A..F) + VLDBench headlines/text
python extract_annotations.py --config vldbench --format parquet --output out_pairs.parquet
# All annotators agree: related + similarity_mean
python extract_annotations.py --config vldbench_unanimous --format csv --output out_unanimous.csv
# Best annotator pair agrees (C+F): related + similarity_mean
python extract_annotations.py --config vldbench_best_pair --format csv --output out_best_pair.csv
# Long multi-annotator rows (IAA only), optional single annotator
python extract_annotations.py --config vldbench_annotations_long --annotator "Annotator C" --format json --output out_c.json
--config is how the script recognizes the subset (same name as the Hugging Face config). Do not pass supplementary_opposing*; load those directly from the Hub.
Rebuild notes
Builders under scripts/ still document the long-form extraction path for VLDBench IDs. Unique-pair / agreement tables were produced from the published freeze; see UNIQUE_PAIR_RELEASE.md.
- Downloads last month
- 22