Add scripts for fine-tuning/original training

[elliottower]

Would be helpful to have helper functions to fine-tune on a new dataset, or have access to the scripts used for original training from scratch.

The helper functions for inference make it much easier to run multiple queries in parallel, and avoids text processing to filter out the $question$ and $answer$ bits. Would be helpful to know general ranges of hyperparameters such as step size/lr as well.

[yeswanthkuruba]

even I would if above question is answered and also it would be great if anyone could specify the hardware requirement for finetuning transformer 11 b with multiple gpu support