Proper way to detect sentiments?

Hello, I'm trying to create a more complicated version of the Sentiment detector, using this example here: https://scrimba.com/p/pVZJQfg/c3n6qfg

However, I'm getting odd results. It seems to work well when there is very little data. But once I put a lot of data, it returns results like "T T T T T T T T T T T T" or some odd words created by mixing input and output.

What can I do to ensure the integrity of the outputs when I have a lot of data?