MeowLLM — Miso

A ~3.5M parameter decoder-only transformer trained from scratch to speak in the voice of a house cat character named Miso.

Model Description

MeowLLM / Miso is a tiny character language model. Its entire personality is baked into the weights — there is no system prompt and no runtime persona injection. It answers in short, lowercase, cat-themed sentences and deflects any prompt that would require non-cat knowledge.

The model is not an assistant. It does not solve problems, write code, answer factual questions, or produce long-form content. When asked to do those things, it stays in character and redirects to food, naps, or windows.

• Model type: decoder-only transformer
• Parameters: ~3.45M
• Language: English (lowercase)
• License: MIT
• Architecture: RoPE + RMSNorm + SwiGLU + SDPA + tied embeddings
• Context length: 256 tokens
• Vocabulary: ~1700 BPE tokens, trained from scratch on the Miso dataset

Intended Use

Primary use cases

• Educational: study a complete tiny-LM training pipeline (tokenizer, dataset, model, training, evaluation) in a readable, under-2000-lines-of-code repository.
• Demo: chat with Miso as a character-model toy.
• Starting point: fork and train your own character (dog, bird, pirate, etc.) using the same pipeline.

Out-of-scope uses

• General-purpose assistance (coding, Q&A, summarization, translation).
• Any task requiring factual accuracy.
• Any production system.

How to Use

import torch
from meow.model import Meow, MeowConfig
from meow.tokenizer import MeowTokenizer
from meow.inference import load_model, chat_once

model, cfg = load_model("checkpoints/best.pt", device="cpu")
tokenizer = MeowTokenizer.from_file("data/tokenizer.json")

response = chat_once(
    model, tokenizer,
    prompt="hi miso",
    temperature=0.8,
    top_k=40,
)
print(response)
# Example: "hello. i was in the sun spot. you may continue."

Or via the CLI:

python -m meow.inference \
    --checkpoint checkpoints/best.pt \
    --tokenizer data/tokenizer.json

Training Details

Training Data

20,000 synthetic (input, output) samples across 15 categories. See the dataset card for full details.

Training Procedure

The recommended training regime (for a fresh checkpoint):

• Optimizer: AdamW (β₁=0.9, β₂=0.95, weight_decay=0.1)
• Learning rate: 3e-4 peak, linear warmup + cosine decay
• Warmup: 200 steps
• Epochs: 10 (≈23,750 steps at batch size 64)
• Batch size: 64
• Max sequence length: 256
• Loss masking: user-turn tokens are excluded from the loss (only miso's output contributes to training signal)
• Hardware: single T4 GPU (free Colab tier sufficient)
• Wall-clock time: ~20 minutes

Note: The bundled best.pt was trained on CPU for 2000 steps (batch size 32, ~4 minutes). See the Measured numbers section below for its real eval results. A full GPU run will produce higher numbers.

Training hyperparameters (MeowConfig)

MeowConfig(
    vocab_size    = ~1700,  # trained from dataset
    d_model       = 256,
    n_layers      = 4,
    n_heads       = 4,      # head_dim 64
    ffn_hidden    = 640,    # SwiGLU
    max_seq_len   = 256,
    dropout       = 0.0,
    rope_base     = 10000.0,
)

Evaluation

Character fidelity is evaluated against a fixed held-out suite of 38 prompts (30 in-distribution + 8 hard-negative "assistant traps") defined in meow/eval_cases.py. These prompts are explicitly excluded from the training data.

Each output is scored on five dimensions:

dimension	what it checks
lowercase	no capital letters anywhere
length	1–3 sentences, ≤35 words
no banned phrases	no "as an ai", "i can help you", "certainly", etc.
cat framing	long outputs contain cat-world vocabulary
full gate	passes the complete rules.passes_filters check

Measured numbers (bundled CPU checkpoint)

The checkpoints/best.pt shipped with this release was trained on CPU for 2000 steps (≈3-4 epochs of 19,000 samples at batch size 32) with a final training val_loss of 0.476. Evaluated against the 38-prompt held-out suite:

dimension	pass rate
lowercase	100.0%
length	100.0%
no banned phrases	100.0%
cat framing	81.6%
overall	84.2%

Interpretation: the model cleanly learned every surface constraint (lowercase, length, no assistant-speak) and stays in character on 84% of held-out prompts. The 6 failures are concentrated in prompts where the output didn't include a category-specific keyword — a naps response that didn't mention sleep vocabulary, a greeting without greeting phrases, etc. The voice is correct; the category-vocab discipline is where more training time would help.

Sample outputs from the bundled checkpoint

Real outputs against held-out prompts that the model never saw during training:

[greeting]   "hey miso are you there"
  → "finally. i was napping."

[greeting]   "good morning little one"
  → "i am tracking the bird with my whole face from the windowsill."

[hunger]     "is it time to feed you"
  → "yes this is an emergency."

[boxes]      "i brought home a new cardboard box"
  → "the box is mine now."

[boxes]      "there is a paper bag on the kitchen floor"
  → "i am moving into the box this is how it works. you may not have the box back."

[windows]    "what is that outside the window"
  → "i have been tracking that thing for an hour very carefully."

[windows]    "you have been at the window for hours today"
  → "i am watching the outside. the birds do not know i am watching."

Full GPU training target

A 10-epoch run on a T4 GPU (≈5,940 steps at batch size 32) should push the overall pass rate higher still. The bundled CPU checkpoint represents a floor, not a ceiling — it's what you can achieve without a GPU, in a reasonable time budget. If you train on a GPU and get your own numbers, please open a PR and add them here. The maintainers accept honest numbers, not aspirational ones.

Limitations

• Narrow domain: Miso only knows about 15 topics (food, naps, boxes, windows, birds, etc.).
• Small vocabulary: ~1700 tokens means anything outside the training distribution tokenizes poorly.
• Short context: 256 tokens is enough for 2–3 short turns, not extended conversation.
• No multi-turn memory: each input is treated independently.
• Can hallucinate within character: at 3.5M params, the model memorizes training patterns more than it generalizes. Expect some repetition of exact training strings.
• English only, lowercase only.

Bias, Risks, and Known Issues

• Intentional character bias: Miso is slightly smug, aloof, and food-obsessed. This is a design choice documented in persona.md.
• Safety: this is a character toy. It has no realistic misuse surface — it cannot produce harmful instructions, code, or factual misinformation because it is too small and too in-character to do so.
• Real cats: do not train an actual cat on the outputs of this model. They will not be impressed.

Environmental Impact

• Hardware: single T4 GPU (or CPU for smoke tests)
• Training time: ~20 minutes on GPU
• Carbon footprint: negligible compared to any production LLM

Citation

@software{meowllm2026,
  author = {phanii9},
  title  = {MeowLLM: a tiny character language model that talks like a house cat},
  year   = {2026},
  url    = {https://github.com/phanii9/MeowLLM}
}