Summary of Stop! Benchmarking Large Language Models with Sensitivity Testing on Offensive Progressions, by Robert Morabito et al.
STOP! Benchmarking Large Language Models with Sensitivity Testing on Offensive Progressions
by Robert Morabito, Sangmitra Madhusudan, Tyler McDonald, Ali Emami
First submitted to arxiv on: 20 Sep 2024
Categories
- Main: Computation and Language (cs.CL)
- Secondary: Artificial Intelligence (cs.AI); Computers and Society (cs.CY)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary This paper addresses the pressing issue of mitigating explicit and implicit biases in Large Language Models (LLMs). Current methodologies evaluate scenarios in isolation, neglecting the broader context and potential biases within each situation. To tackle this challenge, the authors introduce the Sensitivity Testing on Offensive Progressions (STOP) dataset, comprising 450 offensive progressions with 2,700 unique sentences of varying severity, covering a broad spectrum of demographics and sub-demographics. The STOP dataset is evaluated against several leading closed- and open-source models, including GPT-4, Mixtral, and Llama 3, revealing inconsistent bias detection rates ranging from 19.3% to 69.8%. By aligning models with human judgments on STOP, the authors demonstrate improved model answer rates on sensitive tasks like BBQ, StereoSet, and CrowS-Pairs by up to 191%, while maintaining or improving performance. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper is about making sure language models don’t have biases against certain groups of people. Right now, many researchers test these models in isolation, without thinking about how they might work in different situations. The authors created a special dataset called STOP that has many examples of offensive language, covering all sorts of demographics and sub-groups. They tested several popular language models on this dataset and found that even the best ones don’t always detect biases correctly. By making these models match human judgments on the STOP dataset, the researchers showed that they can perform better on tasks like recognizing offensive language. |
Keywords
» Artificial intelligence » Gpt » Llama
