vast / README.md
sileod's picture
Record native stance-label quality finding
2b4dd28 verified
|
Raw History Blame Contribute Delete
1.74 kB
---
dataset_info:
features:
- name: sentence1
dtype: string
- name: sentence2
dtype: string
- name: labels
dtype: string
- name: metadata
dtype: string
splits:
- name: train
num_bytes: 8770069
num_examples: 11164
- name: validation
num_bytes: 1080702
num_examples: 1390
- name: test
num_bytes: 1584881
num_examples: 2007
download_size: 1875437
dataset_size: 11435652
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
- split: validation
path: data/validation-*
- split: test
path: data/test-*
license: other
source_datasets: []
---
# vast
VAST native human stance annotations, using original comments and topics.
Excludes type_idx=4 synthetic neutral pairs. Keeps native train/dev/test,
complete original comments, annotation IDs and article IDs. Original data:
https://github.com/emilyallaway/zero-shot-stance/tree/master/data/VAST.
Dataset license is unspecified; no code license is applied to the comments.
Original data: [original distribution](https://github.com/emilyallaway/zero-shot-stance). Repackaged as parquet for [tasksource](https://github.com/sileod/tasksource) by [scripts/repackage_dataset/stance.py](https://github.com/sileod/tasksource/blob/main/scripts/repackage_dataset/stance.py#L56).
[Conversion provenance](provenance.json): source revisions, hashes, configuration and exclusions.
## Quality status
Not selected for Tasksource training: coarse review found native training row `18506` explicitly endorsing humans being omnivores but labeled against topic `omnivore` (ListTopic/type 3). Labels are preserved, not repaired. Topic/stance alignment requires further validation before training use.