---
dataset_info:
  features:
  - name: uid
    dtype: string
  - name: premise
    dtype: string
  - name: hypothesis
    dtype: string
  - name: label
    dtype: int64
  - name: reason
    dtype: string
  - name: idx
    dtype: int64
  - name: predicted_label
    dtype: int64
  - name: explanation
    dtype: string
  splits:
  - name: train
    num_bytes: 25898135.29584466
    num_examples: 14648
  download_size: 11550000
  dataset_size: 25898135.29584466
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
---
Original dataset is ANLI, from
Nie, Y., Williams, A., Dinan, E., Bansal, M., Weston, J., & Kiela, D. (2020). Adversarial NLI: A New Benchmark for Natural Language Understanding. Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, 4885–4901. https://doi.org/10.18653/v1/2020.acl-main.441

This version is generated with examples that were missed with a Qwen/Qwen2.5-7B-Instruct model in an initial sweep.