DAMO-ConvAI/diana

Diana

The PyTorch implementation of paper Domain Incremental Lifelong Learning in an Open World (ACL 2023)

Requirements

cd DomainIncrementalLL
pip install -r r.txt

QA tasks

Data preparation

Download datasets, unzip and place in /data_process

QA Dataset

Tokenize plain text datasets:

cd downstream
python create_l2p.py --model_name_or_path t5-base --dataset_name null --output_dir null

Train and evaluate on 4 GPUs

bash ./diana.sh ./ll 8 48 

ablations: No task prompt & No meta prompt

To evaluate ablation without task prompts:

bash ./dianawotask.sh ./ll 8 48

To evaluate ablation without meta prompts:

bash ./dianawometa.sh ./ll 8 48

DecaNLP tasks

Dataset

extract plain texts from raw files:

cd downstreamdeca
python extract.py

tokenize plain texts:

python toknize.py

Run the training and evaluation script:

bash ./diana.sh ./ll 8 48