A RNN (Recurrent Neural Network) is a type of AI model that you can train from scratch that is very good at making text! In this demo, you can train, import, export, and use RNNs in your browser!
Warning: It may not work on older devices that don't support JS or are just generally laggy
It uses vanilla JS to train and run the models, which should give the best performance!
Training data: What the AI learns off
Epochs: How many times the AI goes over the training data (the more epochs, the smarter it is! but it takes longer to train)
Train model: Start training. It will train over the model unless reset everything is pressed to reset the model
When the model is training, you will see Epoch __/__ | Loss: ___. The Epoch section shows you how many epochs you have done, and how many you have left. The loss is not too important unless you know what you are doing. Most of us don't!
Export model: Export the entire model to share with your friends!
Import model: Import a model from your friends
Reset everything: Resets the entire model
Context: The "prompt" for the model. The AI will continue whatever you put here
Output token length: Sets the length (in characters) of what the ai can say. eg having output token length set to 1000 would make the AI able to say 1000 characters
Generate text: Makes the model run using context.
If the example from the button looks like garbage, train it more! Write more questions and answers!
If the AI says <eot>, it will stop the message. This is useful if you have a training prompt like "What is 1+1? 2! <eot> What is 2+2? 4! <eot>"
Everyone! Even if you have a terrible computer, you can still use this program! It even works on iphones!
I always disliked teachable machine because they never added a text model, which I thought was the most important thing ever for some reason.
Training data
Context