P-tuning-v2

Why we included this project

P-tuning v2 is a useful option if you want close to full fine-tuning accuracy while updating only a small set of continuous prompt vectors. The method adds trainable prompts to every layer of a frozen pretrained transformer, so you can adapt a model to a new task without the memory and compute cost of classical fine-tuning. The repo comes with runnable scripts and hyper-parameter search tooling for a few standard benchmarks like SuperGLUE and SQuAD, along with sequence tagging, plus a separate module for neural text retrieval. That makes it straightforward to reproduce the reported numbers before trying the approach on your own data. Because the backbone stays frozen, this is a realistic path on modest GPU setups, and the code doubles as a clear reference for comparing deep prompt tuning with prefix tuning.