Parameter-efficient Few-shot Learning based on Prompting

1. Requirements

torch==1.10.1
transformers=4.21.2
sentencepiece==0.1.97
scikit-learn==1.1.2
datasets==2.4.0

2. Preprocessing

Preprocessing few-shot learning data

First, download the raw dataset for building the k-shot few-shot dataset.

bash script/download_dataset.sh
python3 data_util/generate_k_shot_data.py --k k \
                                          --data_dir data_dir \
                                          --output_dir output_dir

Now you can build the features of few-shot dataset for each tasks.

bash script/preprocessing/preprocessing_finetuning_data.sh

3. Training

bash script/few_shot.prompting.training.sh
bash script/conventional_tuning.full_params.training.sh
bash script/conventional_tuning.freeze.training.sh

4. Aggregate Results

bash script/aggregate_results.sh

Contact

judepark@jbnu.ac.kr