StackLlama: fixed RL training and added args

[mnoukhov]

added steps argument and break to respect max training epochs added more PPOConfig args to script args
removed llama tokenizer hacks
black + isort

switched to LlamaTokenizer from AutoTokenizer added return_token_type_ids=False to pipeline kwargs because

• the latter loads LlamaTokenizerFast
• LlamaTokenizerFast will output token_type_ids see LLaMATokenizerFast works abnormally transformers#23818 and 🚨🚨 🚨🚨 [Tokenizer] attemp to fix add_token issues🚨🚨 🚨🚨  transformers#23909
• these token_type_ids cause an error in our reward model pipeline, namely TypeError: LlamaForSequenceClassification.forward() got an unexpected keyword argument 'token_type_ids'

[ArthurZucker]

You can use return_token_type_ids=False as a quick fix to prevent them from being passed

[HuggingFaceDocBuilderDev]

The documentation is not available anymore as the PR was closed or merged.

[younesbelkada]

THank you so much for your contribution! Could you just run the styling checks? After that we should be good for merging

make style && make quality

[mnoukhov]

Fixed style and quality and Switched back to Autotokenizer, thanks for the tip @ArthurZucker

[younesbelkada]

Thanks so much!