C4
Benchmarks
Task NameDataset NameSOTA ResultTrendResults
C4 Llama 2 & 3 (test)
15.022-7 Score
14
C4
3.86Eval Loss
12
C4
13.9C4 Score
12
C4 Message length b=64 v1 (test)
99.69BA
12
C4 Message length b=32 v1 (test)
100Bit Accuracy
12
C4
342Latency (rho=0.8)
12
C4
96.97TP@FP=1%
12
C4 LLaMA-1.3B (val)
13.13Perplexity
12
C4 250 tokens
96.7TPR @ FPR 0.1%
12
C4 150 tokens
0.822TPR @ FPR 0.1%
12
C4
10.9Perplexity (PPL)
12
C4 v1 (test)
6.63Perplexity
11
C4-en LLaMA-1B, 13.1B tokens
14.05Perplexity (PPL)
11
C4 (calibration set)
0KL Divergence (Top 10)
11
C4
100True Positive Rate (TPR)
10
C4 v1 (val)
11.07Perplexity
10
C4
16.22Score at 30K Tokens
10
C4 200 tokens (1,000 human-written samples)
0FPR
10
C4 v1.0 (val)
15.42Perplexity (C4)
9
C4 Llama 3 320M pre-training (val)
25.92Validation Perplexity
9
C4 Token Replacement Attack, T5-XXL (test)
0.97ROC-AUC
9
C4 Token Replacement Attack T5-Large (test)
0.998ROC-AUC
9
C4 Clean Text
1ROC-AUC
9
C4
77.12Accuracy
9
C4 1,000 human-written samples 200 tokens
0.1FPR
9