Controllable Text Generation on Yelp (test)
10.3Perplexity (PPL)FUDGE
Evaluation Results
| Method | Links | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| FUDGElambda=100, base_model=GPT22022.08 | 10.3 | — | — | — | — | 77 | -2.97 | 27.2 | 0.05 | — | |
| FUDGEBackbone=GPT2-large, Approach=Plug-and-play2022.08 | 10.3 | — | — | — | — | 77 | — | 27.2 | — | 77 | |
| GPT2-FTtraining=Fine-tuned2022.08 | 10.6 | — | — | — | — | 88 | -11.31 | 23.8 | 0.049 | — | |
| GPT2-FTBackbone=GPT2-large, Supervision=Fully-supervised2022.08 | 10.6 | — | — | — | — | 98 | — | 23.8 | — | 98 | |
| PPLM2022.08 | 11.8 | — | — | — | — | 86 | -4.68 | 31 | 0.102 | — | |
| PPLMBackbone=GPT2-large, Approach=Plug-and-play2022.08 | 11.8 | — | — | — | — | 86 | — | 31 | — | 86 | |
| PROMPT GATINGCategory=PET (Parameter efficient tuning) w/o JT2022.12 | 21.77 | 0.42 | 84.8 | — | 79.91 | — | — | — | — | — | |
| GPT2-LargeTraining=DP-SGD, epsilon=1, Teacher=None2024.03 | 26.84 | — | — | — | — | — | — | — | — | — | |
| LATENTOPS2022.08 | 30.4 | — | — | — | — | 99 | — | 13 | 0.807 | — | |
| LATENTOPSBackbone=GPT2-large, Approach=Latent space / VAE2022.08 | 30.4 | — | — | — | — | 99 | — | 13 | — | 99 | |
| TAILORCategory=PET (Parameter efficient tuning) w/o JT2022.12 | 40.29 | 0.39 | 80.68 | — | 74.7 | — | — | — | — | — | |
| PROMPT-TUNINGCategory=PET (Parameter efficient tuning) w/o JT2022.12 | 40.89 | 0.42 | 48.29 | — | 48.2 | — | — | — | — | — | |
| DistilDPModel=DistilGPT2, Training=DistilDP, epsilon=2, Teacher=GPT2-Large2024.03 | 44.15 | — | — | — | — | — | — | — | — | — | |
| DistilGPT2 (DPKD)Training=DPKD, epsilon=2, Teacher=GPT2-Large2024.03 | 46.16 | — | — | — | — | — | — | — | — | — | |
| DistilGPT2 (DP-SGD)Training=DP-SGD, epsilon=2, Teacher=None2024.03 | 48.12 | — | — | — | — | — | — | — | — | — | |
| DIST. LENSCategory=PET (Parameter efficient tuning) w/ JT (Joint training)2022.12 | 52.59 | 0.26 | 77.47 | — | 72.22 | — | — | — | — | — | |
| DistilGPT2 (DP Syn Data)Training=DP Syn Data, epsilon=2, Teacher=GPT2-Large2024.03 | 61.84 | — | — | — | — | — | — | — | — | — | |
| DistilGPT2 (Zero-shot)Training=Zero-shot, epsilon=0, Teacher=None2024.03 | 65.46 | — | — | — | — | — | — | — | — | — | |
| PREFIX-TUNINGCategory=PET (Parameter efficient tuning) w/o JT2022.12 | 147.47 | 0.31 | 47.53 | — | 58.32 | — | — | — | — | — | |
| GEDICategory=DTR (Decoding-time regulation)2022.12 | 616.92 | 0.75 | 99.47 | — | 75.41 | — | — | — | — | — |