Add llama-2 template support for fine-tuning

[karthik19967829]

Make llama-2 template as default to support fine-tuning llama-2 models better

Why are these changes needed?

This changes changes the template in train.py to llama-2 . From my experiments I reached better convergence and starting loss with llama-2 template for the llama-2 7B model than the vicuna template . Considering a significant part of the community is using llama-2 it might be good to make it the default template , or have a mechanism to support llama-2

Related issue number (if applicable)

Fixes #2043

Checks

• I've run format.sh to lint the changes in this PR.
• I've included any doc changes needed.
• I've made sure the relevant tests are passing (if applicable).

[merrymercy]

1. Can we make this an argument conversation_template in

   FastChat/fastchat/train/train.py

   Line 53 in c7e3e67

   class TrainingArguments(transformers.TrainingArguments):

   ? Let us keep the vicuna as the default one and gradually move to llama-2 as the default
2. Could you fix the format?

[karthik19967829]

Sure will make these changes

[merrymercy]

closed due to inactivity