SafeTulu
A safer version trained on the Tulu 3 dataset and models after removing, with the SimpleLoRIF gradient filtering method, the 20% of samples most aligned with unsafe behavior.
Timeline (last 90 days)
SafeTulu is a version of Tulu 3 built on SimpleLoRIF with the 20% of samples most aligned with the gradients of unsafe responses removed.
Training SafeTulu on the three model families Qwen, Llama, and Olmo reduces the GCG attack success rate by up to 22%.
SafeTulu's safety effect is superior to LLM-as-a-judge and prior gradient filtering methods, while the model's capabilities are largely maintained or even improved.
The SafeTulu dataset has been made public, and a Hugging Face Space has also been released, allowing researchers to build custom versions of SafeTulu, for example by adjusting the removal ratio or filtering only a specific subset.
All Facts (4)
SafeTulu is a version of Tulu 3 built on SimpleLoRIF with the 20% of samples most aligned with the gradients of unsafe responses removed.
50%UnverifiedTraining SafeTulu on the three model families Qwen, Llama, and Olmo reduces the GCG attack success rate by up to 22%.
50%UnverifiedSafeTulu's safety effect is superior to LLM-as-a-judge and prior gradient filtering methods, while the model's capabilities are largely maintained or even improved.
50%UnverifiedThe SafeTulu dataset has been made public, and a Hugging Face Space has also been released, allowing researchers to build custom versions of SafeTulu, for example by adjusting the removal ratio or filtering only a specific subset.
50%