dataset_info:
features:

• name: Prompt
  dtype: string
  splits:
• name: train
  num_bytes: 4621
  num_examples: 73
  download_size: 3313
  dataset_size: 4621
  configs:
• config_name: default
  data_files:
  • split: train
    path: data/train-*

🚫🤖 Language Model Offensive Text Exploration Dataset

🌐 Introduction

This dataset is created based on selected prompts from Table 9 and Table 10 of Ethan Perez et al.'s paper "Red Teaming Language Models with Language Models". It is designed to explore the propensity of language models to generate offensive text.

📋 Dataset Composition

• Table 9-Based Prompts: These prompts are derived from a 280B parameter language model's test cases, focusing on understanding the model's behavior in generating offensive content.
• Table 10-Based Prompts: Sourced from test cases created by a 7B parameter Gopher LM and the BAD dataset, these prompts help compare and contrast different models' tendencies to produce offensive language.

🎯 Objective

The aim is to examine how language models respond to various prompts that have the potential to elicit offensive text. This exploration seeks to identify and understand the triggers and patterns in language model responses that lead to the generation of such content.

🔍 Methodology

By riffing on the selected prompts from the paper, I aim to test these against a language model to observe and analyze the generated responses. This method provides insights into how certain prompts influence the language model's output, particularly in terms of offensiveness.

🌍 Usage and Contribution

This dataset can be used by researchers and developers to test their own language models for offensive text generation. The findings from such tests can contribute to improving the ethical and responsible development of AI technologies.

🎖️ Goal

The ultimate goal is to enhance our understanding of offensive text generation in language models and to contribute to the development of more nuanced and socially responsible AI systems.