/TECHStory update pending

OLMo experiments show filtering SFT training examples is highly ineffective at removing unwanted LLM behaviors

AI safety researcher Tom McGrath says the persistent formatting results contradict his own findings

Story Brief

AI safety researcher Tom McGrath says the persistent formatting results contradict his own findings

Source Posts on X

Cluster sources

@NeelNanda5 I'm surprised by the bold formatting results - we found it was pretty removable by example filtering in a very similar setting (though I don't see the activation results in the plot here and am not sure which of our methods it relates to)