Fix eval batch size, add Dockerfile, improve logging, remove unused code, freeze requirements

[michal2409]

No description provided.

[itayhubara]

Ok, most of it looks great. Thank you, it was exactly what I planned to do today (post-submission). I have a few questions before merging:

1. what is the difference between paged_adm_w and adamw_torch
2. you changed the default value of Lora alpha to 32 - in your experiments you used 16 right?
3. It seems like you kept the monkey patch on the training_steps - I thought that after fixing the eval_steps it is unrequited.

[michal2409]

1. The paged version is more optimized for memory, but regular adamw works fine
2. I thought we decided to use alpha=32
   
3. Right, I will remove it.

[itayhubara]

ok I understand, we can keep alpha as another hyper parameter. I'll merge