Pogo LM — 4.8M
Pogo Spark is a 4,805,280-parameter decoder-only language model trained from
scratch on an M1 Pro. It is a narrow, original cartoon-repair character model,
not a general-purpose assistant.
This repository contains the selected MLX inference checkpoint from character
training step 300, plus its exact tokenizer, config, checkpoint metadata, and
the fixed behavior prompts used during the local run.
Files
- checkpoint.npz — MLX weights
- checkpoint.json — model geometry and checkpoint integrity metadata
- tokenizer.json — 4,096-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
training/inference code from
DDDD-433/pogo-lm:
1git clone https://github.com/DDDD-433/pogo-lm
2cd pogo-lm
3uv sync --python 3.12
4
5# Download the five files from this Hub repo into ./model/
6uv run pogo-generate \
7 --config model/config.json \
8 --tokenizer model/tokenizer.json \
9 --checkpoint model/checkpoint.npz \
10 --prompt '<|bos|> <|user|> Can you help me fix a wobbly paper rocket? <|assistant|>' \
11 --temperature 0 --top-k 1
Training and limitations
The base stage ran for 12,000 steps on TinyStories, PersonaChat, and original
Pogo data. The selected character checkpoint passed 8/8 fixed local behavior
checks under greedy decoding. That is a regression check, not a safety
certification or a general benchmark.
The corpus policy deliberately excludes named-franchise character data and
scraped scripts. Source details and licenses are documented in the GitHub
repository. The base corpus includes TinyStories, which is licensed under
CDLA-Sharing-1.0; review upstream terms before redistributing the weights.
