Share your thoughts, 1 month free Claude Pro on usSee more
WorkDL logo mark

Llama-3

Benchmarks

Task NameDataset NameSOTA ResultTrend
RefusalLlama-3-8B n≈200
ASR0
42
Jailbreak Attack TransferabilityLlama-3-8b-Instruct finetuned variants v1 (test)
TSR51.2
16
LLM Inference PerformanceLLaMA-3 8B
TTFT (ms)56.03
12
Matrix Multiplication LatencyLlama-3 70B
Kernel Latency (µs)293.82
8
Matrix Multiplication LatencyLlama-3 8B
Kernel-level latency (µs)152.69
8
Watermark DetectionLlama-3-8B Translate perturbation, 30 tokens 1.0 (test)
Mean P0.13
6
Watermark Detection RobustnessLlama-3-8B GPT-4o Paraphrase, 150 Tokens
Mean P0.26
6
Watermark Detection RobustnessLlama-3-8B GPT-4o Paraphrase, 30 Tokens
Mean P29
6
Watermark Detection RobustnessLlama-3-8B Swap 50%, 30 Tokens
Mean P25
6
LLM JailbreakingLlama-3-8B-Instruct
SRF1
4
Adversarial AttackLlama-3-70B successful attacks
Unique Queries Count1,321
3
Adversarial Attack Diversity AnalysisLlama-3-70B
Average Attack Similarity35.2
3
Showing 12 of 12 rows