marola-gpt 0.0.1

A small language model that marola trains from random weights on its own corpus of sourced ocean and marine-safety notes. An educational model: it shows how a GPT learns a small domain from scratch. It is not an assistant, and its text is not advice about the sea.

Architecture	nanoGPT (Andrej Karpathy, MIT), character level: 4 layers, 4 heads, width 128, context 128
Parameters	826,240
Vocabulary	87 characters
Training data	marola-corpus v0.1.0
Trained	600 iterations; best validation loss 2.4854
Code	marola-dev/marola-ml at 6b035325e579; nanoGPT at 3adf61e154c3

Use it

Inside a checkout of marola-dev/marola-ml: put model.pt at sea/nanogpt/out-marola-gpt/ckpt.pt and meta.pkl at sea/nanogpt/data/marola_corpus/meta.pkl, then just marola-gpt-sample.

Licence

MIT (LICENSE, Copyright (c) 2026 marola.dev). The code marola-gpt was trained with that came from nanoGPT is under nanoGPT's own MIT License (Copyright (c) 2022 Andrej Karpathy).

Credits

Andrej Karpathy, for nanoGPT. Fabio Akita, whose microgpt gave marola the idea of a GPT trained from scratch on one machine.