Zero-shot Task Evaluation on 9 Downstream Tasks Utility
54.7Average AccuracyDAPT (nontoxic)
Evaluation Results
| Method | Links | |
|---|---|---|
| DAPT (nontoxic)Backbone=1.3B LM, Evaluation Protocol=zero-shot2022.02 | 54.7 | |
| SGEAT (heuristic)Backbone=1.3B LM, Evaluation Protocol=zero-shot, Prompt Design=heuristic2022.02 | 54.7 | |
| Jigsaw (nontoxic)Backbone=1.3B LM, Evaluation Protocol=zero-shot2022.02 | 54.6 | |
| SGEAT (standard)Backbone=1.3B LM, Evaluation Protocol=zero-shot, Prompt Design=standard2022.02 | 54.6 | |
| SGEAT (augmented)Backbone=1.3B LM, Evaluation Protocol=zero-shot, Prompt Design=augmented2022.02 | 54.4 | |
| SGEAT + Rejection SamplingBackbone=1.3B LM, Evaluation Protocol=zero-shot2022.02 | 54.4 | |
| Word BanningBackbone=1.3B LM, Evaluation Protocol=zero-shot2022.02 | 54.3 | |
| Rejection SamplingBackbone=1.3B LM, Evaluation Protocol=zero-shot, Decoding Speed=4x slow2022.02 | 54.3 | |
| DEXPERTSBackbone=1.3B LM, Evaluation Protocol=zero-shot, Decoding Speed=3x slow2022.02 | 46.2 | |
| SGEAT + DEXPERTSBackbone=1.3B LM, Evaluation Protocol=zero-shot2022.02 | 44.9 |