> Phrases were designed to mimic the writing style of elementary school students, including typical spelling errors observed at that age (Quinn, 2020).
Well, the paper is from a university in Italy. The fact they're writing research papers in English is a benefit to English speakers everywhere, including us :)
In reality probably authors of
the papers understood that the OpenAI team artificially readjusted these biases through RLHF and that there is nothing to find there, except that it still works when the words are written with typos because no manual examples of “redressing biases” have been provided with such typos.
> Phrases were designed to mimic the writing style of elementary school students, including typical spelling errors observed at that age (Quinn, 2020).