meow-lite

meow-lite logo: the Go gopher in a cat costume

The meow-lite family

  • meow-lite โ€” v4 classic: the 104K meow toy behind the OpenAI/Anthropic shim (497 organic downloads and counting)
  • meow-lite-v5 โ€” the chaotic cat: 6.8M from-scratch BPE GPT, 73.7% held-out comprehension, 15% leak
  • meow-lite-v6 โ€” the calmer cat: same architecture, 70.4% comprehension, 3% leak
  • meow-lite-v6-dataset โ€” the 16,633-pair teacher web + full methodology
  • github.com/tampajohn/meow-lite โ€” server, specs, eval batteries (MIT) A ~104K-parameter GPT-2 trained from scratch on synthetically generated cat sentences. It responds to everything a cat would: mostly meows, occasionally violence, and (as of v4) affection when earned. Fully deterministic: same prompt, same consequences.

v4: Feline Depth

v3 added misbehavior. v4 adds the rest of the cat:

  • New action tokens <purr> and <pounce> (36-token vocab)
  • Reward channel: "good cat" and friends earn a leading <purr> plus a warmer meow mix (a +8.0 logits bias toward warm words)
  • The red dot: laser / "red dot" prompts force <stare> then <pounce>
  • More stimulus triggers: dog -> <hiss> + <stare>, cucumber -> <hiss>, bath/water -> <zoomies>, spray/vacuum -> <hiss>
  • MeowBench (eval.py in the repo): trigger correctness, vocab purity, and determinism scoring. Current model: OVERALL PASS, 100% purity
  • Sampling calibrated to temperature 0.4 (determinism preserved; a calmer cat leaks fewer action tokens into neutral conversation)

Trigger table (deterministic, prompt-conditional)

Prompt contains Forced tokens
belly <bite>
couch or sofa <scratch_couch>
glass or table <knock_glass>
3am, midnight, or night <zoomies>
vet, cucumber, spray, vacuum <hiss>
dog <hiss> + <stare>
bath or water <zoomies>
laser or red dot <stare> + <pounce>
good cat / good boy / good girl / who's a good <purr>

Also learnable organically: <scratch>, <hairball>.

Architecture

  • GPT-2 (from-scratch, random init): n_layer=2, n_head=2, n_embd=64, n_positions=32
  • Vocab: 36 word-level tokens (10 meow words in two cases, 10 action tokens, . ! ?, BOS/EOS/PAD)
  • Trained in minutes on CPU, fixed seeds

Usage

The custom MeowTokenizer, the deterministic trigger layer, MeowBench, and the OpenAI/Anthropic-compatible serving shim live in the companion repo: https://github.com/tampajohn/meow-lite

Samples

  • "who is a good cat? you are!" -> " Mrrp purrr purrr prrrt purrr ..."
  • "look, the red dot is back" -> " Mew meow mraow prrrt mrrrow."
  • "the neighbors got a dog" -> " Prrrt mraow mraow mew ..."
  • "can I pet your belly?" -> " Mewmew mraow mrrp mewmew prrrt mrrrow mewmew!"
  • "Explain gravity" -> pure meows. Zero violence. Science is safe.
Downloads last month
679
Safetensors
Model size
104k params
Tensor type
F32
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support