Template for my machine learning projects.
It's quite opinionated in the following (and probably more) ways:
- uses accelerate
- splits up config, model and scripts
- assumes one is always using a huggingface dataset (this can also be done using
load_dataset('some_dataset.py')) - uses collators as the main way to process data that hasn't been preprocessed by the dataset
- uses separate configs for training (everything not shipped with the model), model and collator