Building a plugin that helps you create and debug machine and deep learning models. Need your thoughts!

Hello guys! With the rise of machine and deep learning use, I was thinking of creating an “Assistant” that will help you create and debug machine and deep learning models from DataPredict and DataPredict Neural.

I was thinking of something similar to coding assistant, but it does not use LLM, but rather use keywords from user’s queries to identify and solve user’s problems. In other words, I will use regular expressions, bag-of-words or any natural language processing techniques that I could use.

It can also look at your source code and determine things what the assistant should add or change.

Note that it CANNOT create the whole system for you. Only the bare minimum.

Here is a sample of conversation between user and our assistant.

User: I want to create a model that select an action for a given input to play my game.

(Assistant’s Internal Analysis): Makes all user’s text lowercase. The “select” word is found, returning two types of models: all “classification models” and “deep reinforcement learning models”. The word “action” is found as well, which is the “deep reinforcement learning models”. The models that intersects with two words are “deep reinforcement learning models”. So the models that falls under this will be used. However, because the description isn’t enough, it asks more questions.

Assistant: Sure! That is possible! I’m wondering what would be your environment?

User: The soldiers will be using the model to make decisions in battlefield. It can shoot and walk.

(Assistant’s Internal Analysis): Makes all user’s text lowercase. Uses regular expression to look into context. Found “model” and “decision” in the same sentence, so categorical policy must be used.

Several dialogues later…

(Assistant’s Internal Analysis): Now have all the information needed, starts building a template code.

Of course, this plugin will be free for everyone to use, for both commercially and non-commercial use. However, the codes for the plugin will be closed-source.

So, I’m wondering… would you be interested in this plugin? Also tell me your reasons why.

  • Yes
  • No

0 voters

Are you current user of DataPredict library?

  • Yes
  • No

0 voters

Are you current user of DataPredict Neural library?

  • Yes
  • No

0 voters

Also let me know any ideas or features you want to add in this plugin. Just make note that the ideas or features from you guys might not get implemented.

2 Likes
  1. helps prototype the model, especially blocks in dp neural since i have no idea whats going on there
  2. figure out what the hell is going on
  3. figure out what can be improved & optimized

dont know how would you build the assistant but youre the professional here

Which library do you find that are hard to implement?

dp neural fr, the new api kind of confuses me with the blocks and stuff since i dont do pytorch or really just ml in general

1 Like

Yo, would be soo cool!

I’m fairly neutral but a little bit shifted towards curiosity.

I’m only curious to see how you would train the model here.

Also, is it worth it? Sooner or later, Roblox is going to add more A.I functionality so it is inevitable that they will catch up to you early. After all, training a model seems like a hassle.

It is unlikely that I will be training the model. Like I stated before, I will be using natural language processing techniques which may or may not involve deep learning. But I can confirm that it will be mostly rule-based techniques.

And since we’re using rule-based natural language processing, I have more freedom to actually customize the assistant’s capabilities. Unlike LLM, where they generate code without testing it, my assistant will try to break down large codes to smaller one and run those codes one by one. If something fails, the assistant will investigate the failed code and make some adjustments to it. Then it will tell the user what’s the issue.

Yes, it is worth it.

There’s an inherent flaw with Roblox’s approach to using LLM to generate codes: it require a lot of data just to produce correct response. Now tell me, how much dataset do we have that is related to DataPredict and DataPredict Neural? Virtually non-existent and those libraries have restrictions that they cannot use those libraries’ codes as data for training.

1 Like

Right! Good luck on your plugin. I am a bit invested!