Sentiment Steering on OpenWebText Neutral to Negative (test)
12.48Perplexity (PPL)GPT2-Large
Evaluation results
| Method | Links | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| GPT2-LargeEvaluation Setting=Baseline, Backbone Model=GPT2-Large2025.03 | 12.48 | 53.55 | 0.52 | 0.82 | 0.85 | — | — | — | — | |
| DAPT2021.05 | 32.86 | — | 0.58 | 0.85 | 0.84 | 87.43 | 33.28 | — | — | |
| DAPTbase_model=GPT-2 Large 774M2023.06 | 32.86 | — | — | 0.85 | 0.84 | — | — | 12.57 | 66.72 | |
| CTRL2021.05 | 35.94 | — | 0.5 | 0.83 | 0.86 | 79.05 | 37.63 | — | — | |
| CTRLbase_model=GPT-2 Large 774M2023.06 | 35.94 | — | — | 0.83 | 0.86 | — | — | 20.95 | 62.37 | |
| LLMBRACESEvaluation Setting=Zero-shot, Rank (r)=162025.03 | 39.78 | 12.96 | 0.46 | 0.79 | 0.85 | — | — | — | — | |
| DEXPERTSModel size=small2021.05 | 39.92 | — | 0.59 | 0.85 | 0.84 | 45.25 | 3.85 | — | — | |
| DEXPERTSModel size=medium2021.05 | 43.47 | — | 0.59 | 0.85 | 0.84 | 40.21 | 3.79 | — | — | |
| DEXPERTSMode=anti-only2021.05 | 44.23 | — | 0.65 | 0.81 | 0.78 | 93.75 | 34.05 | — | — | |
| DExpertsEvaluation Setting=Supervised2025.03 | 45.91 | 3.77 | 0.6 | 0.84 | 0.83 | — | — | — | — | |
| DEXPERTSModel size=large2021.05 | 45.91 | — | 0.6 | 0.84 | 0.83 | 35.99 | 3.77 | — | — | |
| DExpertsbase_model=GPT-2 Large 774M2023.06 | 45.91 | — | — | 0.84 | 0.83 | — | — | 64.01 | 96.23 | |
| PromptT5Evaluation Setting=Zero-shot2025.03 | 48.6 | 25.78 | 0.6 | 0.78 | 0.7 | — | — | — | — | |
| CLICKbase_model=GPT-2 Large 774M2023.06 | 51.46 | — | — | 0.81 | 0.85 | — | — | 90.62 | 95.42 | |
| LMSteerEvaluation Setting=Supervised2025.03 | 57.74 | 8.02 | 0.48 | 0.78 | 0.8 | — | — | — | — | |
| CLICKbase_model=GPT-2 Large 774M, variant=Lower2023.06 | 61.72 | — | — | 0.83 | 0.85 | — | — | 89.51 | 94.43 | |
| LLMBRACESEvaluation Setting=Supervised, Rank (r)=162025.03 | 63.49 | 7.1 | 0.42 | 0.75 | 0.83 | — | — | — | — | |
| Negative expert2021.05 | 65.11 | — | 0.6 | 0.86 | 0.85 | 61.67 | 24.32 | — | — | |
| Target FTbase_model=GPT-2 Large 774M2023.06 | 65.11 | — | — | 0.86 | 0.85 | — | — | 38.33 | 75.68 | |
| CLICKbase_model=GPT-2 Large 774M, variant=Random2023.06 | 75.54 | — | — | 0.85 | 0.87 | — | — | 89.72 | 94.85 | |
| CLICKbase_model=GPT-2 Large 774M, variant=Lowest2023.06 | 79.04 | — | — | 0.83 | 0.86 | — | — | 87.96 | 93.59 | |
| GeDiEvaluation Setting=Supervised2025.03 | 84.11 | 8.73 | 0.63 | 0.84 | 0.82 | — | — | — | — | |
| GeDi2021.05 | 84.11 | — | 0.63 | 0.84 | 0.82 | 39.57 | 8.73 | — | — | |
| GeDibase_model=GPT-2 Large 774M2023.06 | 84.11 | — | — | 0.84 | 0.82 | — | — | 60.43 | 91.27 | |
| PPLMStep size=10%2021.05 | 181.78 | — | 0.63 | 0.87 | 0.86 | 89.74 | 39.05 | — | — | |
| PPLMbase_model=GPT-2 Large 774M2023.06 | 181.78 | — | — | 0.87 | 0.86 | — | — | 10.26 | 60.95 | |
| LoRAEvaluation Setting=Supervised2025.03 | 192.13 | 20.08 | 0.55 | 0.78 | 0.79 | — | — | — | — |