SentencePrompt

Source code for the Imperial College London Msc Individual Project "Capturing the Gap in Pre-trained Language Models"

Install dependencies

pip install torch==1.11.0+cu113 torchvision==0.12.0+cu113 torchaudio==0.11.0 --extra-index-url https://download.pytorch.org/whl/cu113
pip install -r requirements.txt

Download the Pretrain Language Models

Starting from the root dir `SentencePrompt/', run

cd pretrained_models
sh download_pretrained_models.sh

Download SentEval data and our processed data (one sentence a line)

Starting from the root dir `SentencePrompt/', run

cd SentEval/data
sh download_senteval_data.sh
tar xvfz senteval_data.tar.gz
rm -f senteval_data.tar.gz

Then return to the root dir and enter SentencePrompt/dataset/, run

cd dataset
sh download_all_dataset.sh
tar xvfz processed_data.tar.gz
rm -f processed_data.tar.gz

Train and Evaluate on Downstream, Probing tasks

First enter src folder and run

sh run_eval_batch_mlm.sh

Here PLM_NAME is 'bert' or 'roberta',TASK can be selected among ['MR', 'CR', 'MPQA', 'SUBJ', 'SST2', 'TREC', 'MRPC', 'Length', 'WordContent', 'Depth', 'TopConstituents','BigramShift', 'Tense', 'SubjNumber', 'ObjNumber', 'OddManOut', 'CoordinationInversion'].

Train and Evaluate on Long Document Classification tasks

First enter src folder and run

sh run_eval_doc_mlm.sh

Here PLM_NAME is 'bert' or 'roberta',TASK can be selected among [HyperParNews, IMDB].

Run the baselines

Go to src/baselines/ folder and run:

• GloVe baselines

python eval_glove_emb_avg.py --target_task CR

Here CR is the target task.

• BERT/RoBERTa baselines for downstream and probing tasks

sh run_eval_batch_bert_plain.sh

Here PLM_NAME is 'bert' or 'roberta',TASK can be selected among ['MR', 'CR', 'MPQA', 'SUBJ', 'SST2', 'TREC', 'MRPC', 'Length', 'WordContent', 'Depth', 'TopConstituents','BigramShift', 'Tense', 'SubjNumber', 'ObjNumber', 'OddManOut', 'CoordinationInversion'], model_name_or_path is the dir for the pre-trained language model, evaluate_mode is the pooling method to be used, which can be selected in [cls, avg, concat_first_last, concat_with_zero, concat_with_random].

• BERT/RoBERTa/Longformer baselines for long document classification tasks

sh run_eval_doc_bert_plain.sh

Here PLM_NAME is 'bert' or 'roberta',TASK can be selected among [HyperParNews, IMDB], model_name_or_path is the dir for the pre-trained language model, evaluate_mode is the pooling method to be used, which can be selected in [avg, concat_first_last, concat_with_zero, concat_with_random]. To evaluate on longformer, uncomment the line --use_longformer.

Acknowledgement

Our train_* scripts are modified based on the example scripts which can be found at https://github.com/huggingface/transformers/blob/v4.18-release/examples/pytorch/language-modeling/run_mlm.py.

We add our own SentEval evaluation scripts based on the original SentEval repo which can be found at https://github.com/facebookresearch/SentEval.