DatologyAI develops gradient method to spot AI training data that weakens safety

DatologyAI's SimpleLoRIF compares the gradients of training examples with those of unsafe responses, looking for examples that could push a model toward unsafe behavior even when their text appears benign. The team used a Qwen 3 0.6B model to score Tulu 3 examples and removed the 20% with the closest gradients to create SafeTulu. The filtered dataset and a tool for making custom versions are available.