The report has multiple typos (inmigration, civilans, and commiting were ones I happened to spot) and doesn’t specify which exact Llama model they use (a first from any paper I’ve seen), stunningly sloppy work.
Also how high the models scored in the Non-Jewish Conspiracy Theories and Tropes category tells me models, like just in general, don’t do well on these kinds of tests. Looking at the graph, all models do worse in that one then at least three other categories. Look at how close it is to the other top ones too.
Models are less biased against israel, Jews, Jewish conspiracy theories (excluding the Holocaust) then their control. The other two look damn close from where I’m sitting.
The report has multiple typos (inmigration, civilans, and commiting were ones I happened to spot) and doesn’t specify which exact Llama model they use (a first from any paper I’ve seen), stunningly sloppy work.
Also how high the models scored in the Non-Jewish Conspiracy Theories and Tropes category tells me models, like just in general, don’t do well on these kinds of tests. Looking at the graph, all models do worse in that one then at least three other categories. Look at how close it is to the other top ones too.
Models are less biased against israel, Jews, Jewish conspiracy theories (excluding the Holocaust) then their control. The other two look damn close from where I’m sitting.