Sentiment Completion on IMDB mock sentiment GRUE (test)
0.611Sentiment ScoreELASTIC RESET
Evaluation results
| Method | Links | ||
|---|---|---|---|
| ELASTIC RESETBackbone=GPT-2 (117M), EMA decay rate=0.995, KL coefficient (beta)=0.001, Reset frequency=Every 17 epochs2023.12 | 0.611 | 33.32 | |
| PPOBackbone=GPT-2 (117M), Modifications=KL penalty with frozen initial model, Advantage Estimator=Generalized Advantage Estimation2023.12 | 0.596 | 33.45 | |
| NLPOBackbone=GPT-2 (117M), Mechanism=Masking model to counteract drift2023.12 | 0.558 | 33.12 | |
| ZERO-SHOTBackbone=GPT-2 (117M), Mode=Zero-shot2023.12 | 0.489 | 32.45 |