8 fine tuning

[philswatton]

UPDATED: A PR to resolve #8, now open for review.

Big picture:

• Implements model fine-tuning
• Implements a partial config structure for the project, with a view towards minimal changes being required to cover the full project pipeline
• Implements Baskerville array jobs\
• Implements wandb usage

The code here:

• Adds functions for loading model, tokenizer, and trainer
• Adds a base config class, along with experiment, data, and model config classes
• The idea is to use a modular setup for configs, with one config per experiment component (data, model, forgetting techniques) to allow us to easily change different components of the code without needing to change all components
  • I anticipate being able to add forget-specific fine-tuning details to model hyperparam configs
• A function that given a fairly simple top level experiment config, spawns several single experiment configs (and, optionally, an associated baskerville array script)
• Example experiment (both normal and top-level), data, model, and hyperparam configs
• scripts/README containing instructions for fine-tuning
• Some additional dependencies
• Unit tests

[jack89roberts]

Few initial thoughts/comments

[jack89roberts]

^ mostly some small refactoring/docs comments. I haven't tried running it but looks good to me. Probably some upcoming headaches thinking about adding in forgetting nicely.