hunt3rx99/meowllm-miso

MeowLLM / Miso Dataset

Training data for MeowLLM — a ~3.5M parameter character language model that speaks as a house cat named Miso.

Dataset Summary

20,000 single-turn (input, output) pairs across 15 topical categories plus a hard-negative "assistant trap" deflection category. All outputs are in a single consistent character voice (lowercase, short, cat-themed, no assistant phrases).

The dataset is produced by a slot-based compositional template generator with strict per-category filtering and de-duplication. No human annotation — everything is deterministic from hand-written slot banks plus a random seed.

Supported Tasks

• Text generation: training a tiny character model that produces responses in a consistent voice.
• Character fidelity evaluation: the held-out prompt suite in meow/eval_cases.py is explicitly excluded from the training split.

Languages

English (lowercase only).

Dataset Structure

Data Fields

Each sample is a JSON object with:

field	type	description
input	str	Human prompt (things a person might say to a cat)
output	str	Miso's in-character response
category	str	One of 15 topical category names
source	str	"template" or "llm" (augmented)

Data Splits

split	samples
train	19,000
val	1,000

Categories

greeting, hunger, naps, boxes, windows, birds, humans, dogs, vacuum, rain, affection, territory, nonsense_questions, being_picked_up, jealousy

Category balance in the train split is between 5.4% and 7.7% per category (fairly uniform via round-robin generation).

Data Creation

Source Data

All data is synthetically generated from hand-written slot banks in meow/generate_data.py. Each category has:

• inputs: a bank of prompt phrasings
• cores: the main response clause bank
• openers: optional leading phrases
• sensories: optional sensory detail extensions
• redirects: optional secondary clauses

During generation, each sample is assembled by picking one core and optionally attaching other slots according to per-category probabilities. The result is then validated against meow/rules.py before being written.

Filtering

Every generated sample must pass:

1. Strict lowercase — no capital letters anywhere
2. Length: 1–3 sentences, ≤35 words, ≥1 word
3. No banned phrases — whole-phrase matching against a list of ~40 assistant-speak patterns ("as an ai", "i can help you", "certainly", "of course", etc.)
4. Per-category keyword requirement — e.g., hunger outputs must mention food-related words; vacuum outputs must mention hiding or the vacuum itself
5. Cat-framing fallback — long outputs without a category-specific rule must contain general cat vocabulary
6. Deduplication — exact (input, output) duplicates are rejected
7. Eval leakage check — any input matching a held-out eval prompt is rejected

Known Limitations

• The dataset is deliberately narrow. Miso knows about 15 topics and nothing else.
• Vocabulary is small (~1700 BPE tokens) and lowercase-only.
• Responses are short by design (1–3 sentences).
• Outputs are in English and reflect a specific cultural framing of "house cat" (American/European domestic cat).

Personal and Sensitive Information

None. All samples are synthetic and describe a fictional cat's views on food, naps, and boxes.

Considerations for Using the Data

Social Impact

This is a tiny educational dataset for a character model. It has no realistic misuse surface — the model trained on it is too small for open-ended generation and too in-character to be useful as a generic assistant.

Bias

Miso is a slightly smug indoor cat. That bias is intentional and documented in persona.md.

Additional Information

Dataset Curators

phanii9

Licensing

MIT License.

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}
}