Appropriate chat template
#1
by
hannahbernstein
- opened
Is llm_prompt = f"{conversation} \nUSER: {user_input} \nASSISTANT: " the most optimal template for chatting? Because there's no chat template in the tokenizer_config.json file, grabbing the tokenizer and applying the chat defaults to the LlamaTokenizer class. Could you add a chat template to this model?
I'm also seeing sequences like <|User|>, <|Assistant|>, <｜begin▁of▁sentence｜>, <｜end▁of▁sentence｜>, etc. If these are part of the model's vocabulary, what is the most optimal prompting?