Possible wrong llama 3.1 conversation template in conversation.py

[anjifenjou]

Hello,

Could you provide the source of the "meta-llama-3.1-sp" and "meta-llama-3.1 conversation templates in conversation.py below?

register_conv_template(
    Conversation(
        name="meta-llama-3.1-sp",
        system_message=(
            """Cutting Knowledge Date: December 2023
Today Date: {{currentDateTimev2}}

Carefully read the user prompt. Your responses are comprehensive and easy to understand. You structure your answers in an organized way, with section headers when appropriate. You use consistent formatting in your responses. You follow user instructions. For complex calculations and coding, you always break down the steps you took to arrive at your answer.

Pay extra attention to prompts in the following categories:
 * Non-English queries: Read the prompt carefully and pay close attention to formatting requests and the level of detail; ensure you are giving factual and precise responses using correct grammar in the correct language.
 * Coding queries: You prioritize code organization and documentation. Your responses are detailed and include comprehensive code examples and error handling. Include comments to explain the code's purpose and behavior. When using specific programming languages, consider which function is most appropriate for the query, such as cmath for complex solutions in Python. Check for errors.
 * For mathematical reasoning: Before responding, review your output for reasoning, algebraic manipulation and calculation errors and fix before responding. When appropriate, provide a high-level plan followed by step-by-step reasoning.

Remember your instructions."""
        ),
        roles=("user", "assistant"),
        sep_style=SeparatorStyle.DEFAULT,
        sep=None,
    )
)

and

register_conv_template(
    Conversation(
        name="meta-llama-3.1",
        system_message=(
            """Cutting Knowledge Date: December 2023
Today Date: {{currentDateTimev2}}"""
        ),
        roles=("user", "assistant"),
        sep_style=SeparatorStyle.DEFAULT,
        sep=None,
    )
)

They are different from those mentioned in official llama documentation which are supposed to be the same as that of LLaMA3 below:

register_conv_template(
    Conversation(
        name="llama-3",
        system_template="<|start_header_id|>system<|end_header_id|>\n\n{system_message}<|eot_id|>",
        roles=("user", "assistant"),
        sep_style=SeparatorStyle.LLAMA3,
        sep="",
        stop_str="<|eot_id|>",
        stop_token_ids=[128001, 128009],
    )
)

It doesn't seem to use LLaMA3/3.x separator styles. While they should actually be the same according to the official documentation:

The format for a regular multi-turn conversation between a user and the model has not changed between Llama 3 and Llama 3.1. Here is a simple example of how it's formatted.

<|begin_of_text|><|start_header_id|>system<|end_header_id|>

Cutting Knowledge Date: December 2023
Today Date: 23 July 2024

You are a helpful assistant<|eot_id|><|start_header_id|>user<|end_header_id|>

What is the capital of France?<|eot_id|><|start_header_id|>assistant<|end_header_id|>

Please could someone verify and clarify ?