Question regarding additional white spaces in Mistral's chat template

[Aratako]

Hello, I am currently using your chat template to adapt Mistral for system messages. Upon inspecting the content, I noticed that in this repository's chat template for Mistral, there are additional white spaces on the assistant side that are not present in the original Mistral-Instruct chat template. Here is the relevant part from the original chat template:

{% elif message['role'] == 'assistant' %}{{ message['content'] + eos_token}}

Source: https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.2/blob/main/tokenizer_config.json

And here is the corresponding part in this repository's chat template:

{% elif message['role'] == 'assistant' %}
    {{ ' ' + content.strip() + ' ' + eos_token }}

Source: https://github.com/chujiezheng/chat_templates/blob/main/chat_templates/mistral-instruct.jinja

These differences in the chat template result in outputs that slightly differ from the original Mistral-Instruct after applying tokenizer.apply_chat_template(). For example, the README in this repository shows the following example of output:

<s>[INST] This is a system prompt.

This is the first user input. [/INST] This is the first assistant response. </s>[INST] This is the second user input. [/INST]

Using the original chat template, the output should look like this:

<s>[INST] This is a system prompt.

This is the first user input. [/INST]This is the first assistant response.</s>[INST] This is the second user input. [/INST]

Why have spaces been added after [/INST] and before the eos_token?

[chujiezheng]

Hi, in a fix a few days ago, I removed the second white space as it is a mistake.

For the current fixed version, the template is like:

{% elif message['role'] == 'assistant' %}
    {{ ' ' + content.strip() + eos_token }}

The first white space is because I found the generation of Mistral-Instruct usually starts with a white space like Hello, .... Since I have applied .strip() here, the starting white space should be added back.

[Aratako]

Oh sorry, I didn't notice you already fixed it! That's my bad.
And I understood the meaning of first whitespace. Thank you for your quick reply!