---
dataset_info:
  features:
  - name: inputs
    dtype: string
  - name: choices_list
    list: string
  - name: labels
    dtype: int64
  - name: metadata
    dtype: string
  splits:
  - name: train
    num_bytes: 6698998
    num_examples: 8819
  - name: validation
    num_bytes: 1367501
    num_examples: 1303
  - name: test
    num_bytes: 1198681
    num_examples: 1528
  download_size: 6548725
  dataset_size: 9265180
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: validation
    path: data/validation-*
  - split: test
    path: data/test-*
license: cc-by-4.0
source_datasets: []
---

# ukpconvarg

UKPConvArg1Strict native two-argument convincingness judgments (CC BY 4.0).

Original source: https://github.com/UKPLab/acl2016-convincing-arguments.
No official train/test partition: deterministically hash-ranked topics
allocate 12 debates to training and two each to validation and test, keeping
both stance sides and all recurring arguments of a debate together.
These are internal holdouts, not the paper's cross-validation folds.

Original data: [original distribution](https://github.com/UKPLab/acl2016-convincing-arguments). 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#L118).

[Conversion provenance](provenance.json): source revisions, hashes, configuration and exclusions.