Some errors when doing evaluation with hf models

[JiaQiSJTU]

I would like to express my appreciation for this repository, particularly for its robust evaluation component.
However, during re-implementation, I encountered some unexpected behaviors in model generations. I have a few suggestions for improving the source code:

1. In this line, it might be more appropriate to replace temperature=0 with sample=False.

2. For the function generate_completions at this location, consider adding "add_special_tokens=False if args.use_chat_format else True".

3. "Because many tokenizers will treat the word after space differently from the original word alone, to be consistent, we add a space before tokenization and remove it after tokenization." However, when using the llama-3 tokenizer, adding a space before "```" results in it being tokenized as a single token, so nothing should be removed?

[hamishivi]

Thanks for your comments! Actually, we use vllm for generation these days wherever possible, so I think we don't bump into 1 and 2 that often. I'll leave this issue open to track these.