(AGENPARL) – STATE COLLEGE gio 27 aprile 2023
Large language models that use internet files to learn how to respond to user prompts about different countries worldwide repeat biased ideas – both positive and negative – found online. Using positive trigger words, like “hopeful” and “hardworking,” when entering prompts can retrain the models and result in less biased responses
Fonte/Source: https://www.psu.edu/news/research/story/positive-triggering-method-reduces-nationality-bias-large-text-generators