TuneTrack
a small log of fine-tuning experiments
I wanted to actually learn how LLM fine-tuning works, not just read about it. So I took a small open model (Llama 3.2, 3B), trained it with LoRA on a free Colab GPU using a handful of Q&A pairs pulled from one of my own projects, and used MLflow to keep a record of what happened each time I changed a setting.
This page is that record, updated every time I run a new experiment. Nothing here is staged. There are 0 runs logged so far, and each one has a short note explaining what I was testing and what actually happened. Some runs also have a real question I asked the fine-tuned model, with its actual answer, so you can see it working, not just a loss number.