Space Pirate LM — 2.0M

Captain Orra Quill is an original space-pirate character model trained from scratch on an M1 Pro. It has 1,995,200 parameters and a 128-token context window.

This repository contains the selected MLX checkpoint from character-training step 375, plus its matching tokenizer, config, checkpoint metadata, and fixed evaluation prompts.

Files

checkpoint.npz — MLX weights
checkpoint.json — model geometry and checkpoint integrity metadata
tokenizer.json — 2,304-token byte-level BPE
config.json — inference/training geometry
evaluation.jsonl — fixed local behavior prompts

Usage

These are custom MLX weights, not a Transformers or GGUF model. Use the runner from DDDD-433/space-pirate-lm:

git clone https://github.com/DDDD-433/space-pirate-lm
cd space-pirate-lm
uv sync --python 3.12

# Download the five files from this Hub repo into ./model/
uv run pogo-generate \
  --config model/config.json \
  --tokenizer model/tokenizer.json \
  --checkpoint model/checkpoint.npz \
  --prompt '<|bos|> <|user|> Captain, the shuttle engine is coughing blue sparks. What is our next move? <|assistant|>' \
  --temperature 0 --top-k 1

Training and limitations

The base stage ran for 3,000 steps on TinyStories and original Captain Orra dialogue. The selected checkpoint cleared 5/6 stricter fixed checks; a paraphrased map-reading prompt is its known weak point.

This is a small character demo, not a broadly reliable assistant. Captain Orra and the Starling are original characters. The base corpus includes TinyStories under CDLA-Sharing-1.0; review upstream terms before redistributing the weights.

Downloads last month: 219