To conduct prompting experiments, select which directory you intend to use (UMR or morphosyntactic) and then run the files in numerical order (begining ith the file that starts with "1_") and proceeding. Each python script will produce a file or a .out file which should be verified at every step to confirm that results are proceeding as desired.
To conduct fine-tuning experiments, the first step is to create a conda environment using the .yml file (finetuning-env.yml). This step is crucial, if you do not use all of these packages you may not get the same results. Then, select the directory which corresponds to the model and type of fine-tuning you are looking to complete (i.e. if you want to do NSL fine-tuning on NLLB, select NSL_nllb_ft). Then, change the following variables in the .sh script to train your desired model: EXPERIMENT_NAME= TARGET_LANG= SYLLABIFY= MODEL_SIZE= Run the .sh file and your model will be saved to a directory with the experiment name you selected.
Model Checkpoints can be found in this Google Drive folder.
Email me with any questions: emarkle26@amherst.edu.