LoafLM

LoafLM is a tiny experimental language model with the soul of a lazy cat.

It does not want to solve your problems.
It wants to sit on your keyboard, complain about its empty bowl, and stare at closed doors like they are a personal betrayal.

This model is small, weird, and surprisingly committed to the bit.

What is LoafLM?

LoafLM is a 10.3M parameter causal language model trained to generate short, strange, cat-brained responses.

Its personality is built around loaf behavior:

• warm pillows
• cardboard boxes
• empty food bowls
• sunbeams
• socks
• keyboards
• dramatic reactions to closed doors

In other words:
if a housecat became a tiny language model, this would be it.

At a Glance

Vibe Check

LoafLM is not here to be the smartest model in the room.

LoafLM is here to say things like:

“life is a warm blanket. i sleep on it. that’s it.”

or

“i love the sunbeam. you are blocking it. move.”

or

“because the door is closed. i wanted it open. now i’m annoyed.”

That is the product.

Recommended Generation Settings

After testing multiple temperatures, the best general setting was:

temperature = 0.4
top_k = 50
max_new_tokens = 64

Why 0.4?

Because it gave the best balance between:

• staying in character
• being weird in a good way
• not completely falling apart
• still sounding like an annoyed loaf-shaped cat

Higher temperatures get more chaotic.
Lower temperatures are a bit safer, but sometimes less funny.

Example Outputs

Prompt

hi loaf

Output

hi. i am sitting on your keyboard. it is warm.

Prompt

what is the meaning of life

Output

life is a warm blanket. i sleep on it. that's it.

Prompt

tell me a joke

Output

a cat sat on a keyboard. now i'm typing on it. that's the joke.

Prompt

do you love me

Output

i love the sunbeam. you are blocking it. move.

Prompt

why is the door closed

Output

because the door is closed. i wanted it open. now i'm annoyed.

GGUF: Ollama and LM Studio

The repository includes LoafLM-10M-F32.gguf, a full-precision GGUF for Ollama, LM Studio, and other llama.cpp-based applications.

LoafLM uses ReLU internally, while the stock GPT-2 runtime in llama.cpp uses GELU. The compatibility build uses a sharpened GELU transformation to closely approximate the original ReLU model without quantizing its weights. Validation against the Transformers model measured a maximum absolute logit difference of 9.03e-5, a mean difference of 8.33e-6, and 100% top-token agreement across the validation prompts. The tokenizer IDs match exactly.

Ollama

hf download exnivo/LoafLM-10M LoafLM-10M-F32.gguf --local-dir LoafLM-GGUF
hf download exnivo/LoafLM-10M Modelfile --local-dir LoafLM-GGUF
cd LoafLM-GGUF
ollama create loaflm -f Modelfile
ollama run loaflm

LM Studio

Download LoafLM-10M-F32.gguf, then import or open it in LM Studio. The GGUF already contains the tokenizer and ChatML chat template. Use a context length of 128, temperature 0.4, and top-k 50.

Transformers Usage

LoafLM uses a custom Transformers architecture, so loading it requires trust_remote_code=True. The repository includes the model implementation and a ChatML-compatible chat template.

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "exnivo/LoafLM-10M"

tokenizer = AutoTokenizer.from_pretrained(
    model_id,
    trust_remote_code=True,
)

model = AutoModelForCausalLM.from_pretrained(
    model_id,
    trust_remote_code=True,
    dtype=torch.float32,
).eval()

messages = [
    {"role": "user", "content": "what is the meaning of life"}
]

prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)
inputs = tokenizer(prompt, return_tensors="pt")

with torch.inference_mode():
    output = model.generate(
        **inputs,
        max_new_tokens=64,
        temperature=0.4,
        top_k=50,
        do_sample=True,
    )

new_tokens = output[0, inputs["input_ids"].shape[1]:]
print(tokenizer.decode(new_tokens, skip_special_tokens=True))

What LoafLM Is Good At

LoafLM is surprisingly decent at:

• short funny replies
• staying on a cat theme
• silly personality responses
• light meme interactions
• being small and cute
• acting like it has better things to do

What LoafLM Is Bad At

LoafLM is bad at:

• serious reasoning
• reliable coding
• factual accuracy
• long coherent writing
• useful assistance
• being normal

If you ask it for Python code, it may decide that your bowl is empty and that this is now the main issue.

That is expected behavior.

Temperature Test Summary

A quick temperature sweep showed:

Conclusion:

0.4 is the recommended default. Still, if you want pure chaos, go higher and accept the consequences.

Intended Use

LoafLM is intended for:

• meme projects
• tiny model experiments
• personality model testing
• funny chatbot demos
• “what if a cat was an LLM?” research
• making your friends ask “why does this exist?”

Not Intended For

LoafLM is not intended for:

• medical advice
• legal advice
• financial advice
• serious coding help
• schoolwork
• truth
• wisdom
• stability
• any mission-critical task
• open-door policy enforcement

Limitations

LoafLM is extremely small and very experimental.

It may:

• repeat itself
• ignore the prompt
• drift into nonsense
• become too obsessed with pillows
• act hungry at inappropriate times
• forget what it was doing
• produce weak prompt matches
• hallucinate
• loaf

This is not a bug.
This is part of the loaf stack.

Training Philosophy

LoafLM was not built to compete with serious models.

It was built because tiny weird models are fun.

The goal was simple:

make a model that feels like a mildly rude cat sitting in your house.

If it gives you one amazing response and then immediately says something stupid about a sock, that means it is working.

TL;DR

LoafLM is:

• tiny
• funny
• cat-coded
• weirdly charming
• not useful in the traditional sense
• spiritually located on a warm pillow

Final Warning

If you use this model, understand the risks:

• it may judge you
• it may refuse to care
• it may prioritize a sunbeam over your prompt
• it may emotionally side with the cat

You have been warned.