This is stock GPT-2 small (124M), unmodified. It is not a fine-tuned model, and nothing in
the South of Tethys project uses it. It is published here for provenance, and this card
exists so that nobody mistakes it for the thing its name suggests.
What it actually is
An early experiment from the project's first weeks. The intent was to fine-tune a small model
on the South of Tethys canon; what was published was the base checkpoint, before any training
happened. The export script still says so in a comment:
model_name = "gpt2" # Replace with your fine-tuned model if you have one
No replacement ever happened, and it should not now — see below.
Why it was never updated
Because the approach was abandoned, and for a good reason rather than through neglect.
GPT-2 small has no instruction tuning. Given real canon about the Lothal Marsh-Lurker and told
plainly not to invent anything, it produced a description of a man holding a snake. Retrieval
was never the problem — that part works and is a separate system — but a 124M base model cannot
write to a brief, and no amount of prompting fixes that.
The project moved to retrieval-augmented generation against an instruction-tuned model, chosen
at runtime. The live service points at Qwen/Qwen2.5-7B-Instruct and never loads this
checkpoint.
What to use instead
The canon itself, which is the part with any value in it:
Entities — around 470 authored records: species, regions, places, discoveries, people
and vocabulary, in database/ with JSON Schema
and a lint that enforces referential integrity.
Retrieval — a Chroma index over the whole corpus, served behind a small FastAPI service.
/lore answers "what does canon say about this place" from the entities themselves.
Generation — any instruction-tuned model, set by environment variable. The service holds
no opinion about which, and swapping it is a config change.
Should you build on this?
No. It is gpt2, and you should use gpt2 — you will get an identical model with a clearer
name and a proper card. This repository is kept because deleting published artifacts breaks
links and hides history, not because it is useful.
