THUDM/P-tuning-v2

Implements deep prompt tuning by injecting continuous learnable prompts at every transformer layer rather than just the input, significantly reducing trainable parameters while maintaining fine-tuning-level performance. Supports diverse NLP tasks including text classification (SuperGLUE), sequence tagging (NER, SRL), and reading comprehension (SQuAD) across BERT and RoBERTa models. Compatible with Hugging Face Datasets API for streamlined data loading and frozen backbone architectures for parameter-efficient adaptation.

2,077 stars. No commits in the last 6 months.