We require pytorch>=1.2 and transformers=2.3.0 Additional requirements are are in

requirements.txt

We have set up the repo to allow for finetuning on two datasets and testing/ eval on 6 datasets. Combinations of finetune/ test datasets are in the table below:

We run finetuning and eval by updating different shell scripts found in the shell_scripts/ folder of the parent directory. The relevant arguments to update are below:

• TOXIC_DIR: Parent directory where different datasets are read in and where tokenized forms of datasets are cached (assumption is that there are dataset-specific subdirectories)
• TRAIN_DATASET : name of the finetuning dataset to use. Options allowed in current iteration of the repository are:
  • founta
  • civil_identities
• DEV_DATASET : name of the challenge dataset on which the finetuned model is to be scores. Options allowed in current iteration of the repository are:
  • founta
  • civil_identities
  • SBIC
  • bibifi
  • covert_comments
  • twitter_aae
• MODEL_DIR : directory where model checkpoints/results will get output
• do_train / no-do_train: when no_train finetuning will run; when no-do_train only eval will run

We have heavily modified run_toxic.py in order to allow for the new loss functions introduced in src/clf_loss_functions.py

To run these debiasing methods you must first train a shallow model. Do this by calling src/shallow_subsample.py

Take the shallow subsamples created and finetune the model by setting the --debias argument to "shallow". Finetune with the training dataset as the 0.5% portion and evaluate on the remainder. An example for this can be found in shell_scripts/shallow_example.sh

Examples for running debiasing can be found in the shell_scripts folder and should be fairly self-explanatory to run with your own folder structure.

Please see the original readme at https://github.com/XuhuiZhou/Toxic_Debias/blob/main/README.md in order to understand the original intent of the code we've modified.

Our report can be found within the same github Project folder titled NLU_final_paper.pdf