Why does the chat template matter so much?

Local models are trained on a specific token format. Send raw text without the right INST tags or header tokens and the model behaves erratically — ignoring the system prompt or refusing to follow turns. The template aligns your input with training.

How do the three formats differ?

Llama 2 wraps the system prompt in SYS tags inside the first INST block. Llama 3 uses special header tokens with begin_of_text and eot_id markers. Mistral has no system role, so the convention is to prepend the system text to the user message inside INST.

Does Mistral really have no system role?

Mistral Instruct's official template has no dedicated system slot. The widely used convention is to put the system instruction at the start of the first user turn inside the INST block, which this tool does automatically.

Do I need an API key?

No. Formatting happens locally in your browser with no network call. Paste your messages, copy the template, and use it with any runtime.

Should I include the trailing assistant header for Llama 3?

Yes. The empty assistant header at the end is where generation begins. The tool adds it so the model knows to start its reply, which is required for correct behavior.

Llama Prompt Formatter

Llama prompt formatter

Hosted APIs hide the chat template from you, but when you run Llama or Mistral locally you have to format prompts yourself — and each family expects a different token structure. Get it wrong and the model ignores your system prompt or mangles the turns. The Llama prompt formatter wraps a system and user message in the exact template each model family was trained on.

How it works

You pick a model family and enter a system message and a user message. For Llama 2, the tool emits <s>[INST] <<SYS>>…<</SYS>> {user} [/INST]. For Llama 3, it uses the header-token format with <|begin_of_text|>, <|start_header_id|> role markers, <|eot_id|> separators, and a trailing empty assistant header where generation begins. For Mistral, which has no system role, it prepends the system text to the user message inside a single [INST] … [/INST] block. Everything runs locally — no API key, no network call — and you copy the result straight into your runtime.

Tips and notes

Match the template to the checkpoint. A Llama 3 template fed to a Llama 2 model (or vice versa) will degrade output — pick the family that matches your weights.
Keep the special tokens intact. Don’t hand-edit <|eot_id|> or [/INST]; the model relies on them to delimit turns.
Mistral’s system text is a convention. It lives inside the user turn, so keep it concise — there’s no separate slot protecting it.
The trailing assistant header is required for Llama 3. Leaving it off means the model has no signal to start replying.