Data filtering works a lot worse than you would expect

Researchers found that data filtering has little effect on most broad SFT behaviors in LLMs, with the exception of refusal. They used various training data attribution methods to identify and filter out data points, but saw no significant reduction in undesired behaviors.