This is a LoRA fine-tuned Gemma 7B on a QA dataset of paul graham essays. I generated the dataset by scraping all the essays, dividing them by paragraphs, then using OpenAI API to have GPT-4o-mini generate "questions" for each paragraph. The API calls ended up costing 14 cents. I then formatted QAs into a jsonl, divided into a 90/10 train/test split, and trained on an H100 I rented from Lambda. Training took about 15 minutes and cost 50 cents, but I spent another 2 hours debugging and downloading what I trained cuz I'm a noob at this stuff, so total training cost was $5. But I learned a lot! Time to do some testing!