Machine Translation on human-collected unfaithful translation De => En (test)
30.8BLEUTarget-constrained tuning LoRA
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Target-constrained tuning LoRASetting=Supervised, Backbone=LLaMA2-7b-chat, Adapter=LoRA2024.06 | 30.8 | 83.2 | |
| Scheduled Sampling tuning LoRASetting=Supervised, Backbone=LLaMA2-7b-chat, Adapter=LoRA2024.06 | 30.5 | 83.1 | |
| R-Drop tuning LORASetting=Supervised, Backbone=LLaMA2-7b-chat, Adapter=LoRA2024.06 | 30.3 | 83.1 | |
| Target-constrained tuningSetting=Supervised, Backbone=LLaMA2-7b-chat2024.06 | 29.6 | 83 | |
| Vanilla Instruction tuning LoRASetting=Supervised, Backbone=LLaMA2-7b-chat, Adapter=LoRA2024.06 | 29.1 | 82.9 | |
| Vanilla Instruction TuningSetting=Supervised, Backbone=LLaMA2-7b-chat2024.06 | 27.3 | 82.2 | |
| Vanilla FewshotSetting=Supervised, Backbone=LLaMA2-7b-chat2024.06 | 27.1 | 81.7 | |
| Reweight Attention (RA)Setting=Unsupervised, Backbone=LLaMA2-7b-chat2024.06 | 24.5 | 79 | |
| Contrastive Decoding (CD)Setting=Unsupervised, Backbone=LLaMA2-7b-chat2024.06 | 24.2 | 78.7 | |
| Vanilla ZeroshotSetting=Unsupervised, Backbone=LLaMA2-7b-chat2024.06 | 23.2 | 77.8 |