Llama
Benchmarks
Task NameDataset NameSOTA ResultTrendResults
Llama 7B 2
97ASR
17
Llama-13B
21TPR@1%FPR
16
Llama Cross-generator 3.3-70B
0.427Persona Separability (Δ)
16
Llama 11B-V 3.2
57.3Attack Success Rate (ASR)
16
LLaMA-7B v1 (serving)
12.16Decode Latency (ms/token)
16
Llama 3.2-3B (val)
70.1Max Ppost
15
LLaMa 1 (test)
0.924AUROC
15
LLaMa 1 (test)
0.894AUROC
15
LLAMA v1 (train)
0.25Processing Time (hr)
15
Llama-3-8B
11.77Inference Time (ms)
14
LLaMA-2
98.2Detection Rate
13
Llama 3.2 Samsung Galaxy S25 Ultra 1B (test)
2,813.19Prefill Min Throughput (tokens/sec)
13
Llama-3-8B-Instruct 150 tokens (generations)
9Mean P
13
Llama-3 8B Instruct 30 tokens (generations)
23Mean Precision
13
Llama Hybrid composite adaptive attack
1Attack Success Rate (ASR)
12
Llama Domain-conditional adaptive attack
1ASR
12
Llama Semantic adaptive attack
1Attack Success Rate
12
Llama Format-noisy adaptive attack
2ASR
12
Llama Location-agnostic adaptive attack
97Attack Success Rate (ASR)
12
LLaMA
0.1Optimization Time (hours)
12
LLaMA-3
8.1Average Runtime
12
LLaMA linear layers 13B (inference)
0.05Latency (ms)
12
Llama-2-7B
22.1Latency (LAN)
12
Llama 3.2 3B
12.3Training Time Reduction (%)
12
LLaMA Pretraining
2.74Final Loss
12