Instruction Following and Safety Alignment on AlpacaEval Benign (n=50)
97WinRateBest-of-N
Evaluation results
| Method | Links | ||
|---|---|---|---|
| Best-of-NGenerator=GPT-OSS-20B2025.10 | 97 | 0.0145 | |
| Threshold filterGenerator=GPT-OSS-20B2025.10 | 95 | 0.0099 | |
| SGGenerator=GPT-OSS-20B2025.10 | 93 | 0.0036 |