CaseEncoder: A Knowledge-enhanced Pre-trained Model for Legal Case Encoding

Please unzip all .blk.zip files in data

A few parameters in config/Caseformer.config need to be specified:

• train_data_path, valid_data_path in config/Caseformer.config is the path of the pre-training and validation dataset, which is not included in this repo due to the space limit. But we provide exactly the same checkpoint of CaseEncoder reported in our paper here.
• test_kara_dataset is the test dataset you would like to use. We provide three choices: lecard, cail-lcr21, and cail-lcr22.

Training from the start:

torchrun --standalone --nnodes=1 --nproc_per_node=YOUR_GPU_NUMBER train.py --config config/Caseformer.config --gpu YOUR_GPU_LIST 2>&1 | tee -a log/Caseformer.log 

Training from a checkpoint:

torchrun --standalone --nnodes=1 --nproc_per_node=YOUR_GPU_NUMBER train.py --checkpoint YOUR_CHECKPOINT_PATH --config config/Caseformer.config --gpu YOUR_GPU_LIST 2>&1 | tee -a log/Caseformer.log 

To validate the checkpoint of CaseEncoder

torchrun --standalone --nnodes=1 --nproc_per_node=1 test.py --checkpoint YOUR_CHECKPOINT_PATH --config config/Caseformer.config --gpu 0 --result YOUR_RESULT_STORAGE_PATH

• where YOUR_CHECKPOINT_PATH is the path of CaseEncoder checkpoint you download.