Home

Awesome

P-tuning v2

Source codes and data for

An optimized prompt tuning strategy achieving comparable performance to fine-tuning on small/medium-sized models and sequence tagging challenges.

Find our previous version P-tuning v1 for knowledge probing and few-shot SuperGLUE. Your kindly starring our repo can greatly encourage us to work harder :)

You may be also interested in our recent work GLM-130B: An Open Bilingual Pre-trained Model (2022-10-06). It is an open-sourced LLM outperforming GPT-3 175B over various benchmarks. Get model weights, do inference and P-Tuning v2 with only 4 * RTX 3090 or 8 * RTX 2080 Ti FOR FREE!

P-tuning v2 leverages deep prompt tuning, which is to apply continuous prompts for every layer input of the pretrained transformer. Deep prompt tuning increases the capacity of continuous prompts and closes the gap to fine-tuning across various settings, especially for small models and hard tasks.

Thanks @rainatam's joint effort in re-organizing codes for publishing!

Commonly Asked Question

  1. Some readers notice a 'mismatch' in SuperGLUE between P-tuning (v1) and P-tuning v2: This is because in P-tuning's SuperGLUE experiment, for fair comparison to PET, we follow its experimental setting where backbone pre-trained model parameters are jointly tuned with continuous prompt embeddings; while in P-tuning v2, we follow Prefix tuning and Lester et al.'s parameter-efficient setting where backbone pre-trained model parameters are frozen.

Reproduce Tips

Since experiments reported in our paper are all conducted on NVIDIA DGX-A100 servers (which might be difficult to acquire), we reimplement P-tuning v2's results on BERT-large/RoBERTa-large with:

We notice that the best hyper-parameters can be sensitive to your server environment and package version. If you do not have the exact same environment, we highly recommend you to run hyper-parameter search in your environment based on our example hyper-parameter search script in search_script and result collection scripts search.py.

Setup

We conduct our experiment with Anaconda3. If you have installed Anaconda3, then create the environment for P-tuning v2:

conda create -n pt2 python=3.8.5
conda activate pt2

After we setup basic conda environment, install pytorch related packages via:

conda install -n pt2 pytorch==1.7.1 torchvision==0.8.2 torchaudio==0.7.2 cudatoolkit=11.0 -c pytorch

Finally, install other python packages we need:

pip install -r requirements.txt

Data

For SuperGLUE and SQuAD datasets, we download them from the Huggingface Datasets APIs (embedded in our codes).

For sequence tagging (NER, SRL) datasets, we prepare a non-official packup here. After downloading, unzip the packup to the project root. Please use at your own risk.

Training

Run training scripts in run_script (e.g., RoBERTa for RTE):

bash run_script/run_rte_roberta.sh

Implemented Results

Currently we have released our reimplementation on following tasks and datasets. More implementation will be released soon.

Released results on BERT-large

BoolQCOPARTEWiCWSCCoNLL04OntoNotes 5.0CoNLL12
Result74.377.080.175.168.384.586.485.3
Total Epochs10080608080403045
Best Epoch5812305617332443

Released results on RoBERTa-large

BoolQCOPARTEWiCWSCCoNLL03CoNLL04OntoNotes 5.0CoNLL12CoNLL05 WSJCoNLL05 BrownSQuAD 1.1SQuAD 2.0
Results84.092.086.673.764.491.888.490.184.789.483.988.1/94.281.3/84.7
Total Epochs10012010050103080604515-3010
Best Epoch8678653132845593713-249

For other hyper-parameters, please refer to the training scripts. If you can not achieve the reported results at the best epoch, there is probably an environmental mismatch and hyper-parameter search is needed.

Citation

If you find our work useful, please kindly cite our paper:

@article{DBLP:journals/corr/abs-2110-07602,
  author    = {Xiao Liu and
               Kaixuan Ji and
               Yicheng Fu and
               Zhengxiao Du and
               Zhilin Yang and
               Jie Tang},
  title     = {P-Tuning v2: Prompt Tuning Can Be Comparable to Fine-tuning Universally
               Across Scales and Tasks},
  journal   = {CoRR},
  volume    = {abs/2110.07602},
  year      = {2021},
  url       = {https://arxiv.org/abs/2110.07602},
  eprinttype = {arXiv},
  eprint    = {2110.07602},
  timestamp = {Fri, 22 Oct 2021 13:33:09 +0200},
  biburl    = {https://dblp.org/rec/journals/corr/abs-2110-07602.bib},
  bibsource = {dblp computer science bibliography, https://dblp.org}
}