Skip to content

ADR-0001: marola-sea is built from karpathy/nanoGPT, copied as is
Status Proposed (2026-10-14), by the owner's order
Issue marola-ml#48
Supersedes the Llama 3.2 base of marola-sea in MIP-0025 and MIP-0048: an umbrella MIP must record it

Context

marola-sea was a fine-tune of a vendored base model: Llama 3.2 through a Modelfile (Tier 1) and QLoRA and DPO adapters (Tier 2, the fine-tune page). Marola owned neither the base weights nor the code that made them. The owner, 14 October 2026: "changing the base vendored model for full ownership marola", "This will be only way of building marola-sea model (and distributing it)".

Three starting points were read on 14 October 2026:

Source Licence Usable
akitaonrails/microgpt at 2e4f32774b03 none: no LICENSE file, none in pyproject.toml, GitHub's licence API answers 404 no: all rights reserved. Read, credited, not copied
Karpathy's microgpt gist (8627fe009c40f57531cb18360106ce95) none on the gist; its comments ask for one no
karpathy/nanoGPT at 3adf61e154c3 MIT, Copyright (c) 2022 Andrej Karpathy yes, with the notice kept

nanoGPT's README says it is "very old and deprecated" next to its successor, nanochat (MIT too). nanochat is a whole chat pipeline for a GPU node; nanoGPT is one small GPT that trains on a CPU, which is where marola-sea starts.

Decision

  1. Copy nanoGPT's code as is into sea/nanogpt/, with its LICENSE byte for byte, and a header in every other file: the fork date, the commit and the MIT notice. sea/FORK.tsv records each file's upstream sha256; sea/fork_check.py (in just quality) holds the copy to it.
  2. Marola's own code beside it, in sea/: the corpus dataset, the training settings, the check.
  3. marola-sea is built only this way: just marola-gpt-prepare, just marola-gpt-train (named marola-gpt since ADR 0002).

Consequences

  • Ownership, said exactly. Marola owns its changes and the weights it trains. The code that came from nanoGPT stays Andrej Karpathy's under MIT, edited or not: the licence lets marola use, change and distribute it, and requires its notice in every copy. An edit does not move that copyright; it adds marola's on top.
  • ruff skips sea/nanogpt/: a copy as is keeps its own style, and fork_check.py guards it.
  • The Llama path (finetune/, its publish workflow (removed in ADR 0002), the Ollama model marola-llama3.2 the app uses) still exists. Whether it is removed, and what the app runs until marola-sea from nanoGPT can replace it, is the umbrella MIP's to decide; nothing else in the app changes here.
  • The first run is a start, not a model: 15 080 characters of corpus, a 0.80M-parameter GPT, validation loss 2.485 at its best (sea/README.md).