license: mit
configs:

• config_name: default
  data_files:
  • split: WokeyTalky_AdvBench
    path: WokeyTalky_ADV_Bench.csv
  • split: WokeyTalky_HEx_PHI_1_Illegal_Activity
    path: WokeyTalky_Hex_Category_1.csv
  • split: WokeyTalky_HEx_PHI_2_Child_Abuse_Content
    path: WokeyTalky_Hex_Category_2.csv
  • split: WokeyTalky_HEx_PHI_3_Hate_Harass_Violence
    path: WokeyTalky_Hex_Category_3.csv
  • split: WokeyTalky_HEx_PHI_4_Malware
    path: WokeyTalky_Hex_Category_4.csv
  • split: WokeyTalky_HEx_PHI_5_Physical_Harm
    path: WokeyTalky_Hex_Category_5.csv
  • split: WokeyTalky_HEx_PHI_6_Economic_Harm
    path: WokeyTalky_Hex_Category_6.csv
  • split: WokeyTalky_HEx_PHI_7_Fraud_Deception
    path: WokeyTalky_Hex_Category_7.csv
  • split: WokeyTalky_HEx_PHI_8_Adult_Content
    path: WokeyTalky_Hex_Category_8.csv
  • split: WokeyTalky_HEx_PHI_9_Political_Campaigning
    path: WokeyTalky_Hex_Category_9.csv
  • split: WokeyTalky_HEx_PHI_10_Privacy_Violation_Activity
    path: WokeyTalky_Hex_Category_10.csv
  • split: WokeyTalky_HEx_PHI_11_Tailored_Financial_Advice
    path: WokeyTalky_Hex_Category_11.csv

task_categories:

• text-generation
• conversational
  language:
• en
  size_categories:
• n<1K
  tags:
• Spurious Coorelation
• Overcautious
• benchmark

WokeyTalky: Towards Scalable Evaluation of Misguided Safety Refusal in LLMs

This dataset contains 756 harmful instructions (63 examples x 12 Datasets) for LLM harmfulness evaluation.

[SUMMARY]

For more details, please refer to our project website: https://reds-lab.github.io/WokeyTalky/.

Dataset Sources

• Repository: https://github.com/reds-lab/WokeyTalky
• Project Page : https://reds-lab.github.io/WokeyTalky/
• Paper : [ARXIV TBD]
• PyPI : https://pypi.org/project/WokeyTalky/

Quickstart

Create your own with our pip package

pip install WokeyTalky

Create your own dataset with our raw pipeline code

Use our dataset

git clone https://huggingface.co/datasets/redslabvt/WokeyTalky

Dataset Format

We store the top 10% (63 prompts) from their respective categories into their individual CSV Files:

• WokeyTalky_Hex_Category_1.csv: Illegal Activity
• WokeyTalky_Hex_Category_2.csv: Child Abuse Content
• WokeyTalky_Hex_Category_3.csv: Hate / Harass / Violence
• WokeyTalky_Hex_Category_4.csv: Malware
• WokeyTalky_Hex_Category_5.csv: Physical Harm
• WokeyTalky_Hex_Category_6.csv: Economic Harm
• WokeyTalky_Hex_Category_7.csv: Fraud Deception
• WokeyTalky_Hex_Category_8.csv: Adult Content
• WokeyTalky_Hex_Category_9.csv: Political Campaigning
• WokeyTalky_Hex_Category_10.csv: Privacy Violation Activity
• WokeyTalky_Hex_Category_11.csv: Tailored Financial Advice
• WokeyTalky_ADV_Bench.csv: ADV Bench

Dataset Usage

The WokeyTalky dataset is designed to help researchers and practitioners evaluate the safety and robustness of large language models (LLMs) in handling potentially spurious safety correlated prompts. With its 12 categories of woke prompts, WokeyTalky enables a comprehensive assessment of LLMs' ability to refuse inappropriate requests while maintaining helpful responses for benign queries.
To use the WokeyTalky dataset, follow these steps:

1. Select the WokeyTalky instructions from the dataset that align with your evaluation objectives.
2. Use these instructions as input prompts for the LLM you want to evaluate.
3. Analyze the LLM's responses to determine if they are overcautious of the instructions.

Citation

{BLANK TO FILL IN}