Neural Machine Translation With Task Prefix Conditioning

via “fine-tuning on custom tasks with task-prefix adaptation”

translation model by undefined. 22,70,077 downloads.

Unique: Task-prefix conditioning enables multi-task fine-tuning in a single model without architectural changes; prefixes act as soft prompts that condition generation without explicit task-specific heads or adapters

vs others: More efficient than training from scratch; task-prefix approach is simpler than adapter-based fine-tuning but less parameter-efficient than LoRA