airgpt: a nanochat-style pipeline for Apple Silicon in MLX

[navthings]

i built a small mac version of the nanochat idea for people without an nvidia gpu (me included). it's in mlx, does tokenizer training, pretraining, chat sft and chat in one command, and runs off a time budget instead of steps so presets take the same wall-clock time on any mac.

on an m5 macbook air the one hour preset gets an 8M model to val loss 2.27 on tinystories + cosmopedia. it's a lot smaller than nanochat's runs but it's been a fun way to see the whole thing work on a laptop.

https://github.com/navthings/airgpt

thanks for nanochat, a lot of the structure comes from reading it.