This is the repo for ICL CIPHERS.
The dataset folder contains examples files for each dataset used in the paper.
To install the required dependencies
pip install -r requirements.txt
To generate the token frequencies using wikipedia corpus, which will be used in zipfian shuffling
python generate_tf.py --model {model} --hf_token {hf_token}
To accelerate the priority sampling of demonstrations, we first create mappings between all the tokens and demonstrations.
python gen_demo_dict.py --dataset {dataset} --model {model} --hf_token {hf_token}
To run ICL CIPHERS, run icl_all.py.