Chat template

#4
by bartowski - opened

Hey, just wanted to clarify something about the chat template

On the model card you put

GPT4 Correct User: Hello<|end_of_turn|>GPT4 Correct Assistant: Hi<|end_of_turn|>GPT4 Correct User: How are you today?<|end_of_turn|>GPT4 Correct Assistant:

but the chat_template in tokenizer_config.json compiles to:

<|start_header_id|>GPT4 Correct User<|end_header_id|>

{prompt}<|eot_id|><|start_header_id|>GPT4 Correct Assistant<|end_header_id|>

Which is correct?

@bartowski I ran into the problem, that llama-server automatically selected openchat as a chat template. This caused the occurence of multiple <end_of_turn> tokens in the response.

The correct chat template for OpenChat 3.6 seems to be llama3 with the Llama 3 style EOT token: <|eot_id|>

Sign up or log in to comment