Bug fixing and patching

[vietanh125]

In this commit:

• ChatML bug that leads to double <|im_end|> at the end has been fixed.
• Model card (REAME) showing "None dataset" has been fixed.
• Added input prompt printing before training for debugging purposes.

[casper-hansen]

Could you add an example of the prompt format before and after?

[vietanh125]

Sure. For ChatML format, I set the special tokens as follow

special_tokens:
    bos_token: "<s>"
    eos_token: "<|im_end|>"
    pad_token: "</s>"
    unk_token: "<unk>"

tokens:
    - "<|im_start|>"
    - "<|im_end|>"

and it results to an extra <|im_end|> before each user turn like this:

The issue is solved after setting add_eos_token=False on ShareGPTPromptTokenizingStrategy

[winglian]

hi @vietanh125 I can't edit this PR directly, but if you can run the pre-commit hooks, that would be great! pre-commit run --all-files thanks!

[vietanh125]

I ran it, there were some errors in pylint, mypy, and bandit but everything has passed when pushing.

[vietanh125]

I will close this PR and open one with more details about the issues.