GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
lordlebu
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
lordlebu
Model Tree
Input Modalities
Output Modalities
Supported Functionality
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:
python
model_name = "gpt2" # Replace with your fine-tuned model if you have one
No replacement ever happened, and it should not now — see below.
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.
The canon itself, which is the part with any value in it:
database/ with JSON Schema
and a lint that enforces referential integrity./lore answers "what does canon say about this place" from the entities themselves.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.