[Todo] Adapter Implementation

[peradura]

By Next week

• Use Greedy decoding while memory adaptation process. (Disable sampling to immitate the original model's representation)
• Check what kind of dataset is used and how the adaptation process is working on Longformer-Build.
• Refactor the adapter in a way that allows the user to change the dataset and optimizer.
  • Do not use string for type setting method. Allow the optimizer (and etcs) object to be passed directly to the function.
• Use fancy tqdm progress output. (tqdm.auto)

Future work

• Create a unified base class or script that integrates Trainer, Evaluater, and Adapter.
• Combine the Hugging Face Trainer with our own Trainer and make it available as a single interface.

[b-re-w]

@peradura This ISSUE is not for riverLethe. I'll transfer this to the right place.

[b-re-w]

@3GID Are you sure that this Create a unified base class or script that integrates Trainer, Evaluater, and Adapter. is the only Future work that I requested? 🤨