jglo0708 / corruptionflagger

Corruption Flagger

This is a project for fine-tuning a multilingual BERT based classification model for detecing corruption risk flags.

Set up:

1. Run python3 -m venv venv_cf
2. Run source venv_cf/bin/activate
3. Run pip install -r requirements.txt

It is required to have installed CUDA on your device!

To run the script, use python3 run.py PATH_TO_DATA IS_DATA_JSON LABEL_COLUMN TEXT_COLUMN e.g. python3 run.py inputs/json_dir/ True risco txt

Finetune using an external dataset

positional arguments:

• data_path: path to the dataset dir
• data_is_json: is dataset in json format
• label_column: Column with the data label (e.g. fraud/no fraud)
• text_column: Column with the text data (e.g. procurement description, etc.)

optional arguments:

• checkpoint_path: path to save model checkpoints (default: 'checkpoints')
• batch_size: Batch size - change if needed (reduce for lower memory usage) (default: 32)
• n_epochs: Number of epochs to run - change if needed (default: 10)
• max_sequence_len: Max number of tokens to use in BERT model (default: 256)
• bert_architecture: Underlying Bert model (default: "distilbert-base-multilingual-cased")
• learning_rate: Learning rate for the optimiser (default: 2e-5)
• save_transformers_model: Save Transformers model (default: True)
• resume_from_checkpoint: In case you wish to continue training, input the relative path of the .cpt file (default: None)
• run_test: In case you wish to run the model in test mode (default: False)