THUDM/P-tuning-v2
An implementation of deep prompt tuning (P-tuning v2), a parameter-efficient fine-tuning strategy for pretrained transformer models.
Velocity · 7d
+1.2
★ / day
Trend
→steady
star history
P-tuning v2 applies continuous prompts to every layer input of pretrained transformer models, enabling comparable performance to full fine-tuning with fewer trainable parameters. It targets small and medium-sized models as well as hard sequence tagging tasks. The approach was published at ACL 2022 and has been extended to text retrieval tasks in EMNLP 2023 findings.