Question Answering on PubMedQA (B-1, R-L)
9.7BLEU-1KIDGPT2
Evaluation results
| Method | Links | ||
|---|---|---|---|
| KIDGPT2Backbone=GPT2, Implementation=Knowledge trie, Training Strategy=Inference-time optimization2022.04 | 9.7 | 11.5 | |
| Little KIDImplementation=naive (plain list), Training Strategy=Inference-time optimization2022.04 | 5.2 | 6.6 | |
| FiD-T5Model Scale=base, Backbone=T5, Training Strategy=LM FT2022.04 | 4 | 4.4 | |
| FiD-T5Model Scale=large, Backbone=T5, Training Strategy=LM FT2022.04 | 4 | 4.5 | |
| PPLMTraining Strategy=Inference-time optimization (LM FT / -)2022.04 | 3.8 | 4.2 | |
| Diverse DecodingTraining Strategy=Inference-time optimization (LM FT / -)2022.04 | 3.5 | 3.8 |