Please place the Huggingface weight file of the language model to be edited into models/ directory and write config of DAFNet for this model similar to configs/dafnet/llama-7b.yaml. Then add corresponding path into the function get_model_path_editor_config_name in utils/utils.py.
Please run:
python train_dafnet.py --model_name "model_name" --device 0 --extra_device 1 2
where replace model_name with the name of language model you have added into get_model_path_editor_config_name.
Checkpoints will be saved in train_records/dafnet/model_name/train_name/checkpoints/.
You can view training information in train_records/dafnet/model_name/train_name/logs/ through Tensorboard.
In eval.sh, please set -ckpt as the checkpoint path of DAFNet and set -mn as the name of language model to be edited. Then run:
sh eval.sh
You can check results in eval_results/dafnet.
If you want to implement a new language model editor, please inherit the base editor class editors.editor.BaseEditor and base editor config class editors.editor.EditorConfig.