llama3 instruct template will have different outputs depending on system tokens

[Sneakr]

What is the issue?

Issue:
The absense or presence of a system token produces different outputs, based on my findings:
ggerganov/llama.cpp#7062 (comment)

This is even more important for fine tunes on the instruct models as it can break everything.

Official ollama - llama3 template removes tokens if no system message is present:
https://github.com/ollama/ollama/blob/main/docs/modelfile.md
https://ollama.com/library/llama3:instruct/blobs/8ab4849b038c

The corrected template should be:
TEMPLATE """<|start_header_id|>system<|end_header_id|> {{ .System }} <|eot_id|>{{ if .Prompt }} <|start_header_id|>user<|end_header_id|> {{ .Prompt }} <|eot_id|>{{ end }} <|start_header_id|>assistant<|end_header_id|> {{ .Response }} <|eot_id|>"""

Edit:
I've made an official thread and awaiting response:
meta-llama/llama3#203

[Sneakr]

So meta just changed the template page without answering my issue :) Atleast give some credit where credit is due.