All the files relevant to the project, should be placed to data directory. To download them, use the following command:
gdown --folder https://drive.google.com/drive/folders/1x-_NbgxpSgcM-WsjMzWgH-NC6oWleixL
If you get "Permission denied", please use the link above to download data directory and put it to the root directory of the project.
To predict relevant knowledge candidates for each dialog turn, run the following, specifying argument --filename (the name of a .json file from data directory).
python knowledge_selection.py --filename test_focus.json
Before PeFT, reformat the files with functions data2instructions and then configure from utils.py. Please refer to their docstrings for the details.
The script was adapted from the unsloth notebook. To fine-tune the model, run the following command, specifying argument -g: KG for knowledge grounding, PG for persona grounding, or KG_PG for knowledge and persona grounding:
python inference.py -c ZS -g KG_PG -m train
To run prediction on test, run the following command, specifying arguments -c (FT for fine-tuning or ZS for zero-shot), and -g (see fine-tuning section):
python inference.py -c FT -g KG_PG -m predict
To evaluate the quality of outputs, run the following command, specifying argument --filename (the name of a .jsonl file from data directory that has fields output_generated, output, and only for PG predicted_personas).
python eval.py --filename test_focus_KG_PG_generated.jsonl