Made a tool that builds its own training data and improves each cycle by learning from what it got wrong

The basic idea is pretty simple. You give it a few seed prompts. It generates instruction-response pairs, an LLM scores each one, the good ones go into your