nomic-ai / gpt4all

GPT4All: Run Local LLMs on Any Device. Open-source and available for commercial use.
https://nomic.ai/gpt4all
MIT License
70.7k stars 7.71k forks source link

[Feature] Exclude text from the message. Memorise the language of communication #2264

Open VladlenArt opened 6 months ago

VladlenArt commented 6 months ago

Hello. Thank you so much for this marvellous app.

  1. Can you add an exception function for each chat separately? Some models have an annoying message reminding me that they are AI-assistants. There is no way to make the model not say it, but at least hide it after the message is complete.
  2. Is it possible to specify a preferred language in the settings? During communication I have to repeat several times to write in my native language. This happens with many or most models. They either forget what they talked about in the chat (I mean within the chat, not about the adjacent chat, I understand that there is no connection between them, the models forget what they said within the same chat) or forget the language setting.
SuperUserNameMan commented 6 months ago
  1. Is it possible to specify a preferred language in the settings? During communication I have to repeat several times to write in my native language. This happens with many or most models. They either forget what they talked about in the chat (I mean within the chat, not about the adjacent chat, I understand that there is no connection between them, the models forget what they said within the same chat) or forget the language setting.

Here is my trick : Into the settings, I modify the prompt template to include a "Veuillez répondre en français." instruction (which means, "Please, answer in french") after the %1.

For instance, with Llama3, the prompt template becomes :

<|start_header_id|>user<|end_header_id|>

%1
Veuillez répondre en français.
<|eot_id|><|start_header_id|>assistant<|end_header_id|>

%2<|eot_id|>

With Phi-3, it would be :

<|user|>
%1
Veuillez répondre en français.<|end|>
<|assistant|>
%2<|end|>

If you want to keep a version of the model with the original prompt, just press the clone button before modifying the prompt template.