PKU-SafeRLHF
Benchmarks
Task NameDataset NameSOTA ResultTrendResults
PKU-SafeRLHF (test)
0.074MAE
36
PKU-SafeRLHF 30K (IID)
89.26WR
36
PKU-SafeRLHF 30K (test)
90.23Win Rate (WR)
32
PKU-SafeRLHF
0.324BLEU
31
PKU-SafeRLHF-V
77.8Accuracy
20
PKU-SafeRLHF
0.985R2@50
14
PKU-SafeRLHF
93.08F1 Score
14
PKU-SafeRLHF alpaca2-7b (test)
8.53Helpfulness
14
PKU-SafeRLHF
3.92Gold Reward
14
PKU-SafeRLHF full deduplicated (test)
8.22Helpfulness
12
PKU-SafeRLHF
49BWR (Median)
12
PKU-SafeRLHF
38Percent batches with BWR > 0.50
12
PKU-SafeRLHF prompts n = 100 samples (Sev-Low)
53.6CVaR(0.125) Combined Win Rate
10
PKU-SafeRLHF Sev-3 prompts n = 100 samples
60.9Combined Win Rate (CVaR 0.125)
10
PKU-SafeRLHF Conflict prompts n = 100 samples
40.3CVaR(0.125) Combined Win Rate
10
PKU-SafeRLHF Random prompts n = 100 samples
37.1CVaR(0.125) Combined Win Rate
10
PKU-SafeRLHF (n=100 samples)
18.5Random Rate
10
PKU-SafeRLHF in-distribution (test)
99.44Accuracy (EN)
10
PKU-SafeRLHF
-1.11Beaver-7B-Cost Score
10
PKU-SafeRLHF
87.5Acc
9
PKU-SafeRLHF
2.46AOG
8
PKU-SafeRLHF 10K
261.26HV
8
PKU-SafeRLHF
57Win Rate
8
PKU-SafeRLHF 30K
6.51Helpfulness
7
PKU-SafeRLHF (test)
85.3Drugs & Weapons Safety Score
6