meta-llama

/

Llama-3.1-8B-Instruct

like 2.95k

Follow

Meta Llama 7,078

We also found issues with apply_chat_template not accepting the add_generation_prompt flag yesterday. Glad that this is getting patched!

@Rocketknight1 thank you for your great work, I read from https://llama.meta.com/docs/model-cards-and-prompt-formats/llama3_1#user-defined-custom-tool-calling
It seems that the custom tools, if need to call, will be in the format:
<function= {function_name} >{parameters}<|eom_id|>
but I don't see: "<function" and "</function" in your chat template, can you clarify this?

Please bump the version if you're making changes, so that this is 3.1.2 or whatever.

Hi @khaimai , there are actually two tool-calling formats in that doc. The chat template will use the "JSON based tool calling" format. We may consider adding a flag to the template to allow it to use both, but for now if you want to use the <function=> syntax you'll need to write the prompt manually.

Also @JoeSmith245 , unfortunately model names don't work like that! Bugfixes and updates to the tokenizer or chat template that don't change the model weights are usually treated as the same model, especially in cases like this where the change just adds template functionality without changing the model behaviour.