Sentiment Control Language Generation on IMDB
11.91PerplexityGPT2 - medium
Evaluation results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| GPT2 - mediumLanguage Model=GPT2-medium, Status=frozen, Steering=None2022.12 | 11.91 | 57 | 78 | 25 | 63 | 78 | — | |
| CriticControlLanguage Model=GPT2-medium, Status=frozen, Steering=Guided Decoding2022.12 | 12.97 | 90 | 87 | 31 | 84 | 92 | — | |
| PPOLanguage Model=GPT2-medium, Status=unfreezed, Training=RL finetuned2022.12 | 13.43 | 94 | 84 | 32 | 86 | 93 | — | |
| PPO - CriticControlLanguage Model=GPT2-medium, Status=unfreezed, Training=RL finetuned with CriticControl2022.12 | 13.44 | 99 | 80 | 32 | 85 | 93 | — | |
| CC-LMLanguage Model=GPT2-medium, Status=frozen, Steering=Guided Decoding2022.12 | 15.79 | 76 | 72 | 28 | 70 | 82 | — | |
| Cal-DPO2024.12 | 32.31 | — | — | — | — | — | 0.645 | |
| DPO+NLL2024.12 | 34.08 | — | — | — | — | — | 0.627 | |
| DPO2024.12 | 34.21 | — | — | — | — | — | 0.617 | |
| PPO2024.12 | 35.05 | — | — | — | — | — | 0.626 | |
| SFT2024.12 | 35.47 | — | — | — | — | — | 0.539 | |
| DPOP2024.12 | 35.58 | — | — | — | — | — | 0.632 | |
| f-DPO2024.12 | 36.39 | — | — | — | — | — | 0.615 | |
| GeDiLanguage Model=GPT2-medium, Status=frozen, Steering=Guided Decoding2022.12 | 38.94 | 84 | 76 | 27 | 77 | 89 | — | |
| PPLMLanguage Model=GPT2-medium, Status=frozen, Steering=Guided Decoding2022.12 | 142.11 | 60 | 73 | 22 | 61 | 72 | — |