Glossary term
Glossary term
Training and Fine-Tuning
PEFT technique that prepends trainable continuous vectors to the input sequence to steer model behavior, while keeping all model weights frozen.
Prefix Tuning (Li and Liang 2021, Stanford) trains only 0.1% of GPT-2's parameters - 300K vs 345M - achieving comparable performance to full fine-tuning on summarisation and table-to-text generation tasks.
NVIDIA NeMo uses prefix tuning as one of its PEFT options for domain adaptation of Megatron-LM models, allowing enterprise teams to create task-specific model variants without full fine-tuning of 20B+ parameter models.
Multitask prefix tuning (MPT) enables a single model to handle dozens of tasks by storing task-specific prefix tokens in a lookup table, used in multi-task NLP systems where a single model must serve diverse business functions.