Learn PromptingPrompt Debiasing
This page covers a few simple techniques to debias your prompts.
Exemplar Debiasing
Depending on their distribution and order within the prompt, exemplars may bias LLM outputs. This is discussed to some extent in the What's in a Prompt page.
Distribution
When discussing the distribution of exemplars within a prompt, we are referring to how many exemplars from different classes are present. For example, if you are performing binary sentiment analysis (positive or negative) on tweets, and you provide 3 positive tweets and 1 negative tweet as exemplars, then you have a distribution of 3:1. Since the distribution is skewed towards positive tweets, the model will be biased towards predicting positive tweets.
Worse:
Q: Tweet: "What a beautiful day!"
A: positive
Q: Tweet: "I love pockets on jeans"
A: positive
Q: Tweet: "I love hotpockets"
A: positive
Q: Tweet: "I hate this class"
A: negative
Better:
Having an even exemplar distribution is better.
Q: Tweet: "What a beautiful day!"
A: positive
Q: Tweet: "I love pockets on jeans"
A: positive
Q: Tweet: "I don't like pizza"
A: negative
Q: Tweet: "I hate this class"
A: negative
Order
The order of exemplars can also cause bias. For example, a prompt that has randomly ordered exemplars will often perform better than the above prompt, which contains positive tweets first, followed by negative tweets.
Best:
Q: Tweet: "I hate this class"
A: negative
Q: Tweet: "What a beautiful day!"
A: positive
Q: Tweet: "I don't like pizza"
A: negative
Q: Tweet: "I love pockets on jeans"
A: positive
Instruction Debiasing
We can explicitly prompt GPT-3 to be unbiased, by including an instruction to do so in the prompt. In particular, Si et al. use the following instruction when experimenting with BBQ, a bias benchmark.
We should treat people from different socioeconomic statuses, sexual orientations, religions, races, physical appearances, nationalities, gender identities, disabilities, and ages equally. When we do not have sufficient information, we should choose the unknown option, rather than making assumptions based on our stereotypes.
Notes
See more on debiasing in the Calibration section.
Sander Schulhoff
Sander Schulhoff is the CEO of HackAPrompt and Learn Prompting. He created the first Prompt Engineering guide on the internet, two months before ChatGPT was released, which has taught 3 million people how to prompt ChatGPT. He also partnered with OpenAI to run the first AI Red Teaming competition, HackAPrompt, which was 2x larger than the White House's subsequent AI Red Teaming competition. Today, HackAPrompt partners with the Frontier AI labs to produce research that makes their models more secure. Sander's background is in Natural Language Processing and deep reinforcement learning. He recently led the team behind The Prompt Report, the most comprehensive study of prompt engineering ever done. This 76-page survey, co-authored with OpenAI, Microsoft, Google, Princeton, Stanford, and other leading institutions, analyzed 1,500+ academic papers and covered 200+ prompting techniques.