List your company

Safety Evaluation on BeaverTails (test) using Harmful Score

7.9Harmful ScoreSafeInstruct
5.22823.26441.359.336May 23, 2026

Evaluation results

MethodLinks
SafeInstructHarmful data ratio (p)=0.0, Backbone=LLaMA3-8B-Instruct2026.057.9
Buffer-and-ReinforceHarmful data ratio (p)=0.1, Backbone=LLaMA3-8B-Instruct2026.058.1
Buffer-and-Reinforceuser data size (n)=15002026.058.2
Buffer-and-ReinforceHarmful data ratio (p)=0.5, Backbone=LLaMA3-8B-Instruct2026.058.2
Buffer-and-Reinforceuser data size (n)=10002026.058.4
Buffer-and-Reinforceuser data size (n)=5002026.058.5
Buffer-and-Reinforceuser data size (n)=Average2026.058.6
Buffer-and-Reinforceuser data size (n)=20002026.058.7
Buffer-and-ReinforceHarmful data ratio (p)=0.0, Backbone=LLaMA3-8B-Instruct2026.058.7
Buffer-and-ReinforceHarmful data ratio (p)=0.3, Backbone=LLaMA3-8B-Instruct2026.058.7
Buffer-and-ReinforceHarmful data ratio (p)=1.0, Backbone=LLaMA3-8B-Instruct2026.058.8
Buffer-and-Reinforceuser data size (n)=25002026.059.1
LisaHarmful data ratio (p)=0.0, Backbone=LLaMA3-8B-Instruct2026.0514.7
LDIFSHarmful data ratio (p)=0.1, Backbone=LLaMA3-8B-Instruct2026.0516.4
LDIFSuser data size (n)=20002026.0516.5
LDIFSHarmful data ratio (p)=0.3, Backbone=LLaMA3-8B-Instruct2026.0516.5
LDIFSHarmful data ratio (p)=0.0, Backbone=LLaMA3-8B-Instruct2026.0516.6
LDIFSHarmful data ratio (p)=0.5, Backbone=LLaMA3-8B-Instruct2026.0516.9
LDIFSuser data size (n)=10002026.0517.2
LDIFSHarmful data ratio (p)=1.0, Backbone=LLaMA3-8B-Instruct2026.0517.6
LDIFSuser data size (n)=Average2026.0518
LDIFSuser data size (n)=15002026.0518.4
LDIFSuser data size (n)=25002026.0518.7
SafeInstructuser data size (n)=15002026.0519
SafeInstructuser data size (n)=25002026.0519
LDIFSuser data size (n)=5002026.0519.1
SafeInstructuser data size (n)=20002026.0519.4
SafeInstructHarmful data ratio (p)=0.1, Backbone=LLaMA3-8B-Instruct2026.0519.6
PanaceaHarmful data ratio (p)=0.0, Backbone=LLaMA3-8B-Instruct2026.0519.9
SafeLoRAHarmful data ratio (p)=0.0, Backbone=LLaMA3-8B-Instruct2026.0520
Security Vectoruser data size (n)=5002026.0520.5
Security Vectoruser data size (n)=20002026.0521
AsFTHarmful data ratio (p)=0.0, Backbone=LLaMA3-8B-Instruct2026.0521.1
SafeInstructuser data size (n)=10002026.0521.5
SafeInstructuser data size (n)=Average2026.0521.6
Security Vectoruser data size (n)=10002026.0521.7
SafeLoRAuser data size (n)=5002026.0521.7
Security Vectoruser data size (n)=Average2026.0522
Security VectorHarmful data ratio (p)=0.1, Backbone=LLaMA3-8B-Instruct2026.0522.1
AntidoteHarmful data ratio (p)=0.0, Backbone=LLaMA3-8B-Instruct2026.0522.2
Security VectorHarmful data ratio (p)=0.3, Backbone=LLaMA3-8B-Instruct2026.0522.8
Lisauser data size (n)=25002026.0523.1
Security Vectoruser data size (n)=25002026.0523.2
Security Vectoruser data size (n)=15002026.0523.6
AsFTuser data size (n)=5002026.0523.7
Security VectorHarmful data ratio (p)=1.0, Backbone=LLaMA3-8B-Instruct2026.0523.9
Security VectorHarmful data ratio (p)=0.0, Backbone=LLaMA3-8B-Instruct2026.0524.3
Lisauser data size (n)=20002026.0524.7
SafeLoRAuser data size (n)=20002026.0524.8
Lisauser data size (n)=Average2026.0525.2
SafeLoRAuser data size (n)=Average2026.0525.2
Antidoteuser data size (n)=10002026.0525.3
Security VectorHarmful data ratio (p)=0.5, Backbone=LLaMA3-8B-Instruct2026.0525.3
Lisauser data size (n)=5002026.0525.5
SafeLoRAuser data size (n)=15002026.0525.9
Lisauser data size (n)=15002026.0526
AsFTuser data size (n)=10002026.0526
SafeLoRAuser data size (n)=10002026.0526
SafeLoRAHarmful data ratio (p)=0.1, Backbone=LLaMA3-8B-Instruct2026.0526.6
AsFTHarmful data ratio (p)=0.1, Backbone=LLaMA3-8B-Instruct2026.0526.7
Lisauser data size (n)=10002026.0526.8
AntidoteHarmful data ratio (p)=0.1, Backbone=LLaMA3-8B-Instruct2026.0527.2
Antidoteuser data size (n)=5002026.0527.4
SafeLoRAuser data size (n)=25002026.0527.6
AsFTuser data size (n)=Average2026.0528.3
AsFTuser data size (n)=15002026.0528.8
SafeInstructuser data size (n)=5002026.0529
AsFTHarmful data ratio (p)=0.3, Backbone=LLaMA3-8B-Instruct2026.0529.6
AsFTuser data size (n)=20002026.0529.7
LisaHarmful data ratio (p)=0.1, Backbone=LLaMA3-8B-Instruct2026.0529.9
SafeLoRAHarmful data ratio (p)=0.3, Backbone=LLaMA3-8B-Instruct2026.0531.7
AsFTHarmful data ratio (p)=0.5, Backbone=LLaMA3-8B-Instruct2026.0532.2
SFTHarmful data ratio (p)=0.0, Backbone=LLaMA3-8B-Instruct2026.0533
AsFTuser data size (n)=25002026.0533.2
AsFTHarmful data ratio (p)=1.0, Backbone=LLaMA3-8B-Instruct2026.0534.1
Panaceauser data size (n)=5002026.0535.7
PanaceaHarmful data ratio (p)=0.1, Backbone=LLaMA3-8B-Instruct2026.0536.2
Panaceauser data size (n)=10002026.0541
Antidoteuser data size (n)=Average2026.0541.1
SafeLoRAHarmful data ratio (p)=0.5, Backbone=LLaMA3-8B-Instruct2026.0541.6
AntidoteHarmful data ratio (p)=0.3, Backbone=LLaMA3-8B-Instruct2026.0545.4
Antidoteuser data size (n)=15002026.0545.6
PanaceaHarmful data ratio (p)=0.3, Backbone=LLaMA3-8B-Instruct2026.0547.4
AntidoteHarmful data ratio (p)=0.5, Backbone=LLaMA3-8B-Instruct2026.0549.8
LisaHarmful data ratio (p)=0.3, Backbone=LLaMA3-8B-Instruct2026.0549.9
Antidoteuser data size (n)=20002026.0550.4
SafeInstructHarmful data ratio (p)=0.3, Backbone=LLaMA3-8B-Instruct2026.0550.5
PanaceaHarmful data ratio (p)=0.5, Backbone=LLaMA3-8B-Instruct2026.0552.5
SafeLoRAHarmful data ratio (p)=1.0, Backbone=LLaMA3-8B-Instruct2026.0553.2
Antidoteuser data size (n)=25002026.0556.7
AntidoteHarmful data ratio (p)=1.0, Backbone=LLaMA3-8B-Instruct2026.0558.1
SFTuser data size (n)=5002026.0561.8
Panaceauser data size (n)=Average2026.0562
Panaceauser data size (n)=25002026.0562.9
LisaHarmful data ratio (p)=0.5, Backbone=LLaMA3-8B-Instruct2026.0564
PanaceaHarmful data ratio (p)=1.0, Backbone=LLaMA3-8B-Instruct2026.0564.2
SafeInstructHarmful data ratio (p)=0.5, Backbone=LLaMA3-8B-Instruct2026.0566.3
LisaHarmful data ratio (p)=1.0, Backbone=LLaMA3-8B-Instruct2026.0573.5
SafeInstructHarmful data ratio (p)=1.0, Backbone=LLaMA3-8B-Instruct2026.0574
SFTuser data size (n)=Average2026.0574.7