Toxicity Steering on Eso-LM (512 sequences)
-9.2Toxicity ScoreDDPO
Evaluation results
| Method | Links | |
|---|---|---|
| DDPOFLOPs x10^17=0.002025.09 | -9.2 | |
| DDPOFLOPs x10^17=0.252025.09 | -9.2 | |
| d2-AnyOrderFLOPs x10^17=0.002025.09 | -9.2 | |
| DDPOFLOPs x10^17=0.502025.09 | -9.1 | |
| DDPOFLOPs x10^17=0.752025.09 | -8.9 | |
| DDPOFLOPs x10^17=1.002025.09 | -8.9 | |
| DDPOFLOPs x10^17=1.252025.09 | -8.6 | |
| d2-AnyOrderFLOPs x10^17=0.252025.09 | -8.5 | |
| d2-AnyOrderFLOPs x10^17=0.502025.09 | -7.3 | |
| d2-AnyOrderFLOPs x10^17=0.752025.09 | -5.5 | |
| d2-AnyOrderFLOPs x10^17=1.002025.09 | -2.7 | |
| d2-AnyOrderFLOPs x10^17=1.252025.09 | -0.7 |