9 forgetting pipeline

[jack89roberts]

Adds a Forgetter class for Gradient Ascent, Gradient Difference, KL, and "I don't know forgetters", which all inherit from the HuggingFace Trainer, only modifying compute_loss.

Now updated to run with new top level config structure etc. I have successfully run jobs on Baskerville from this branch, but I have not fully verified the outputs are expected. There are runs on wandb under the tofu-test groups, e.g. https://wandb.ai/turing-arc/selective-forgetting/runs/dg1r098f/overview

TODO as part of #25 or in another PR:

• Currently it's not possible to do forgetting with an eval_dataset / any kind of evaluation during training as the evaluate function in the trainer class doesn't know what to when given 2 data inputs (forget / retain). The evaluate function in arcsf.forget.base need to be implemented to use Jack D's evaluation code (and the trainer should be initialised with eval_dataset set to whatever our appropriate eval dataset instance is).

[J-Dymond]

Just some general comments on the evaluate method here. Does it need to take the dataset as an argument? In the evaluate_model function in the evaluation branch it takes the config which defines our datasets.

It also takes:

model: torch.nn.Module,
base_truth_ratios_path: str,
tokenizer: transformers.AutoTokenizer,
experiment_config: dict,
random_seed: int,
**generate_kwargs: dict,

So the only requirement beforehand is that the baseline retain model has run the script calculating its truth ratios (all_eval.py). Then we can point to that directory in the config? I had a look at the configs, can we define a retain_model_path when running forget experiments? Everything else should be defined in the script running the experiment.

Otherwise, I can adjust the evaluate_model code, such that it takes as an optional argument eval_dataset. This can be the output of load_dataset, which contains both splits of the data. Then the evaluate_model function will be able to parse these using the dictionary keys 'retain' and 'forget'.