When type Chinese, it isn't show all the replies

[chrislemke]

When type Chinese, it isn't show all the replies

[llity]

I reinstall the ChatFred workflow downloaded from GitHub using "cft" keywords, but the Chinese content also doesn't be showing completely. By the way, I using the 'cf' keywords, it work fine.

Typing the 'cf' keywords

The English content is correct.

[llity]

Great, thanks for your attention to this matter.

[llity]

Yep, I will track this issue, and help fix it.

[llity]

How to set the maximum tokens for the InstructGPT and ChatGPT, respectively? Thanks.
It needs a bigger maximum tokens value for the InsructGPT model and keeps the ChatGPT model's value as the default.
I change the number of maximum tokens, it works fine. Thanks.

[chrislemke]

Hey @llity

I just released a new version. There you adjust the Maximum tokens for ChatGPT and InstructGPT separately.

[chrislemke]

Hey @llity!
It looks like you are using quite an old version of the workflow. Could you check which one it is?

[chrislemke]

I will investigate and come back to you.

[chrislemke]

Please have a look at the screenshot I posted. I marked the response from the model. Unfortunately, I don't speak Chinese but it looks like the model is exactly returning what is shown.

[llity]

It's Chinese Unicode, but the workflow is not getting the total response result, and the content was broken. Is it need to be modified the value of "maximum tokens"?

Thanks.

I use a tool to convert it.

[chrislemke]

Hm 🤔, you can try changing the Maximum tokens but the default value should be correct.

For the InstructGPT models to work as a chatbot, we must provide a specific response. That is why there is:

def prompt_from_query(query: str) -> str:
    """Creates a suitable prompt for the OpenAI InstructGPT API."""
    return f"Q: {query}\nA:"

We add a Q: in front of the query and a \nA: at the end of it. This may makes problems with non-Latin languages. You could try modifying this in the text_completion.py file.

[llity]

I try your suggestion, but it's not working. I think the InstuctGPT models don't return the answer completely, or the workflow breaks the answer when I type Chinese in it. And if the answer is Chinese should use ".decode('unicode-escape')" to convert the Unicode string.

cf model gets the correct answer.

[chrislemke]

Thanks for the update. If you want you can make some changes and create a pull request if you want to contribute.

[chrislemke]

Hey! Cool that you figured it out. Unfortunately, at the moment it is not possible to have two different Maximum values. But feel free to work on it and open a PR.