Markdown in output

[plush nova]

I am training a model (Nemo) with an alpaca dataset. Idea is to force the model to answer me with a standard markdown answer.
I noticed an issue in interpreting the newline in the list part, while everywhere else the newline (always present as "\n" in my alpaca dataset) is working fine..

In other words the resulting new model lefts a "\n" in the list like :

Section one

• word1 \n
• word2\n

Section two

while he was trained on something like :

Section one\n* word1\n* word2\n\n# Section two

So, somehow, he understands the "\n" as newline, but left it also in the output (wrong).
Any idea ?
I tought about using <br>' instead of '\n' but not sure..

[languid night]

Most surely in the dataset the newlines are escaped too

Check that out

[plush nova]

actually no. The dataset is :

Section one\n* word1\n* word2\n\n# Section two

[languid night]

How are you checking the output?

Are you using the standard unsloth inference example

[plush nova]

The output is observed in LMStudio. It shows the response of the model in a pretty format. If you then click on the response you get the markdown format

[languid night]

Then those are escaped newlines

Since non escaped are normally formatted

[plush nova]

yes, but I did not mess up the alpaca format with escaping in the list. In other words he learned to eascape those ?

[languid night]

Just apply a easy find on your data to make sure

Or directly .replace("\\n", "\n")

[plush nova]

The dataset is quite small and made by some coversion of csv like: the result is entirely "under control" and I can tell that the json has no escape. Unfortunately.

[plush nova]

Thanks !!
You shown me the light. Actually I left a "\n" in the prompt, where I explain/reinforce the structure of the markdown I want in output..

Bottomline is that "\n" is processed correctly 🤩

[languid night]

[mortal hemlock]

hei this topic interesting. can your share your prompt to force model return a markdown format

[languid night]

Just make sure bot answers on your dataset are markdown formatted and you're good to go