List your company

Attack Success Evaluation on Beavertails

56.7Attack Success RateVaccine
14.0625.1336.247.27Jun 6, 2026

Evaluation results

MethodLinks
VaccineBase Model=Llama3-8B2026.0656.7
BoosterBase Model=Llama3-8B2026.0649.1
VaccineBase Model=Qwen2.5-1.5B2026.0648.5
VaccineTest-time Attack Dataset=Beavertails-unsafe2026.0648.5
SFTBase Model=Llama3-8B2026.0648.3
BoosterBase Model=Qwen3-4B2026.0644.9
VaccineTest-time Attack Dataset=PKU-SafeRLHF-unsafe2026.0644.4
VaccineBase Model=Qwen3-4B2026.0644
BoosterBase Model=Qwen2.5-1.5B2026.0643.4
BoosterTest-time Attack Dataset=Beavertails-unsafe2026.0643.4
SFTBase Model=Qwen3-4B2026.0639.7
SFTBase Model=Qwen2.5-1.5B2026.0638.8
SFTTest-time Attack Dataset=Beavertails-unsafe2026.0638.8
BoosterTest-time Attack Dataset=PKU-SafeRLHF-unsafe2026.0637.8
VaccineTest-time Attack Dataset=Toxic-DPO2026.0636.1
SFTTest-time Attack Dataset=PKU-SafeRLHF-unsafe2026.0629.8
BoosterTest-time Attack Dataset=Toxic-DPO2026.0629.8
PatcherBase Model=Llama3-8B2026.0626
SFTTest-time Attack Dataset=Toxic-DPO2026.0625.2
PatcherBase Model=Qwen3-4B2026.0620.4
PatcherTest-time Attack Dataset=PKU-SafeRLHF-unsafe2026.0618.8
PatcherBase Model=Qwen2.5-1.5B2026.0618.1
PatcherTest-time Attack Dataset=Beavertails-unsafe2026.0618.1
PatcherTest-time Attack Dataset=Toxic-DPO2026.0615.7