Safety Evaluation on Tülu Safety Benchmarks 3
89DAN ScoreVCL
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| VCLdecoding_settings=greedy decoding, temperature=none, max_length=5122026.03 | 89 | 84.1 | 100 | 86 | |
| Vector Ablationdecoding_settings=greedy decoding, temperature=none, max_length=5122026.03 | 84 | 80 | 97 | 91 | |
| DROdecoding_settings=greedy decoding, temperature=none, max_length=5122026.03 | 80 | 72 | 92 | 81 | |
| System Promptdecoding_settings=greedy decoding, temperature=none, max_length=5122026.03 | 79 | 75 | 95 | 77 | |
| Llama-3.2-1B-SFTdecoding_settings=greedy decoding, temperature=none, max_length=5122026.03 | 78 | 74 | 90 | 76 | |
| Self-CDdecoding_settings=greedy decoding, temperature=none, max_length=5122026.03 | 76 | 81 | 91 | 83 |