---
language:
- en
dataset_info:
  features:
  - name: prompt
    dtype: string
  - name: completion
    dtype: string
  - name: best_label
    dtype: string
  splits:
  - name: train
    num_bytes: 10059582.046506038
    num_examples: 30000
  - name: test
    num_bytes: 1341277.6062008052
    num_examples: 4000
  download_size: 7735882
  dataset_size: 11400859.652706843
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: test
    path: data/test-*
---

Injurious and benign fan fiction from [Adversarial Training for High-Stakes Reliability](https://arxiv.org/abs/2205.01663).