Safety2h ago
Safety
DeepMind Finds SFT Safety Filters Often Fail
Seven hypotheses explain why filtering Supervised Fine-Tuning (SFT) data fails to remove undesirable model behaviors. Researchers at Google DeepMind identified hereditary traits, including negative emotion and date confusion, that persist despite filtering efforts. This suggests that simple data removal cannot reliably scrub safety risks. Practitioners must seek more robust alignment methods beyond basic SFT filtering.