Characterizing and measuring linguistic dataset drift

Experiments assessing different drift metrics in their ability to predict NLP model performance on novel data. Drift metrics can be computed for any text datasets, and the full experiments (e.g. model fine-tuning) can be run for any sequence classification task. Steps 2a, 2b, and 2c below can be run largely independently from one another, unless otherwise specified (e.g. if only the model-agnostic metrics are desired, then 2b can be omitted). The specific drift outputs from step 2 do not need to be understood in detail, because they will be compiled automatically in step 3a. More detailed usage for each script can be found in the Python script file headers.