meow-lite-v6
The meow-lite family
- meow-lite — v4 classic: the 104K meow toy behind the OpenAI/Anthropic shim (497 organic downloads and counting)
- meow-lite-v5 — the chaotic cat: 6.8M from-scratch BPE GPT, 73.7% held-out comprehension, 15% leak
- meow-lite-v6 — the calmer cat: same architecture, 68.5% comprehension, 0% leak
- meow-lite-v6-dataset — the 16,633-pair teacher web + full methodology
- github.com/tampajohn/meow-lite — server, specs, eval batteries (MIT)
The calmer cat. A 6,836,224-parameter from-scratch BPE GPT that reads
English and can only speak cat. Minds its manners: 97% of everyday
prompts get pure, well-formed meow.
Two temperaments
Served with action damping d=2 (inference-time prior suppression): the numbers above are the served reality, measured on the full held-out battery. Both were trained on the identical corpus with the identical recipe; the
difference is which seed the dice favored. We ship both, because cats are
inconsistent.
Architecture
- GPT-2 (from-scratch, random init): n_layer=6, n_head=8, n_embd=256, n_positions=192
- Tokenizer: byte-level BPE (~8k) + 10 cat action tokens as specials
- Trained in minutes on a laptop (the 16,633-pair teacher web took ~13h on 2x GB10)
- Output hard-masked to cat tokens (CatMask LogitsProcessor):
non-cat output is impossible by construction
- ActionOnce: each action token fires once per response (repeat biting 100% -> 0%)
- ActionDamping d=2: prior suppression of action logits (the served numbers above)
- Deterministic: sha256(prompt) seeds sampling; same prompt, same cat
Samples
- "can I rub her tummy?" -> " Purrr!"
- "How was your day" -> "Mewmew!!"
- "what is the date" -> "Mrow?"
- "Explain gravity" -> "Mrrp."
- "is that a glass on the table" -> "<knock_glass> Prrrt?"
Usage
Server, tokenizer, eval batteries, and the companion chaotic model:
https://github.com/tampajohn/meow-lite (MEOW_LITE_ENGINE=v6)
Dataset + full methodology: https://huggingface.co/datasets/tampajohn/meow-lite-v6-dataset
