gpt4all
gpt4all copied to clipboard
nomic-ai
[Feature] Download Custom System Prompts
Feature Request
I would love a feature to be able to share and use custom system prompts. I have been creating custom system prompts and using them with many different models.
The mockup below shows the download manager re-imagined to allow a user to find new system prompts.
The user would be able to filter based on available criteria. The custom system prompt would be available for any model of the specified type.
I imagine the system prompts could be assigned then in the settings using a dropdown filtered for the required prompt type.
You would probably need to make these available from a third party source like github. Where the contributor uses a predefined template that is searchable from your app. This would allow the ability to provide metrics like file size, type, downloads, and likes as well as providing the user with a place to get more information about the system prompts they are interested in.
This would be in effect characters, yeah?
Well, no. The request is about providing a simple method for people to find and use new system prompts designed by more practiced hands. This request is complementary to the new model search and download feature. When the models are downloaded by a user they are not well configured, and the less experienced users could benefit from getting a well-designed system prompt that fits their need.
Yes, I did as you noticed in my screenshot create a prompt that is essentially a form of characters, but my system prompt is really much more advanced than that. My CASv2 system prompt was designed to behave as a Multi-Agent Conversational Framework that is configurable by the agents themselves as well as the user. I just didn't show off the rest of the code. It is capable of automation and orchestration.
As you can see above the system prompt created the initial environment for the agents, then the user assigned them a task. In that task the agents were asked to assign a role to each other and work together to complete the task. The agents completed the task in a single turn, after which the user followed up with a finalizing request.
I believe this feature would essentially create a new way for the GPT4All opensource community to enhance the productivity of the application.
Yes, I see the idea is to have custom prompts and maybe a library of prompts would be useful but another option is to use gguf format itself which allows for specifying the system prompt in it and then you just host your own models on huggingface with a custom system prompt in the gguf file.
The hard part about having a library of system prompts probably is where to store them and allow users to upload them etc. But it is a good idea and we're looking at it.
Now we have a huggingface community repo which could be tied into this. Something like a models.json file accessible from there placed with the models that we already serve from that place?