I built a local web UI to finetune models on my own text and actually watch the training (works on AMD ROCm)

I wanted to do continued pretraining/finetuning of a local model on my own notes and see what's happening while it trains and do it on my AMD card, since most tools assume CUDA.

llm-training-panel is a local web UI that:

- loads a model from a local dir or a HF id

- shows a live board of per-parameter gradient activity while it trains

- runs base vs trained on the same prompt so you can A/B what changed

- scores recall + verbatim overlap against a probe file

Tested on Linux + AMD ROCm. Results vary with base model/corpus/settings

Feedback welcome.

github.com/limkcreply/llm-training-panel