GPT-3 Bias and Job Ads

When looking at the bias in the different models it would be great to have some sort of baseline or comparison to the real world. So with the data set we acquired for the project so we actually compared like you mentioned real world job ads compared to GPT-3 level job ads. We found that the baseline zero shot job ads were more biased than the job ads that we found online. However then and this is kind of speaking to this we then went ahead and selected some real world job ad examples with low bias and essentially fed it into the model for fine tuning. Then saw that the frequency with which these stereotypical words appeared was decreasing as a consequence.