AIP is a terminal-based training monitor I built while training my own 1.3B-parameter Polish language model from scratch on an RTX 4080.
When you're running a multi-day DeepSpeed training job, you need a lightweight way to track what's happening without leaving the terminal or spinning up a web dashboard. AIP fills that gap — it shows epoch progress, live loss curves rendered as a histogram, per-batch stats, GPU memory usage, ETA to epoch end, and a live log tail, all in a single TUI screen. Built in Go. No browser, no Python dependencies, no Jupyter. Just SSH into your box and see exactly what your model is doing.
The screenshot above is from epoch 15 of 30 — loss sitting around 2.74, GPU holding steady at 1.9GB, roughly 213 hours left on the run. Exactly the kind of visibility you need when you're not babysitting the machine.