You can use a smaller model like GPT-2. You are not going to get ChatGPT performance without a terabyte of VRAM, but if you want to try something locally, GPT-2 exists.

Humans are more complex than the binary positive/negative. You may have negative attributes which your partner compliments with their positive attributes, and you may have positive attributes which your partner compliments with their negative attributes. If both partners are willing to learn from each other, this can lead to massive growth as you each give an example of how to change the other's negative characteristics.

Exactly. Researchers often need to implement things from scratch, so we use more code-heavy solutions to have deep control. Someone putting these techniques into use may prefer just to load up a pre-trained model or run a training script on an existing architecture.