LOLLM

A language model trained entirely from LOLCODE.

HAI 1.4
CAN HAS BRAINZ?
I HAS A model ITZ I IZ BRAINZ'Z LOAD YR "brainz.bin" MKAY
I HAS A out ITZ I IZ BRAINZ'Z SPEEK YR model AN YR "An Ceiling Cat maded " AN YR 200 AN YR 0.8 MKAY
VISIBLE "An Ceiling Cat maded " AN out
I IZ BRAINZ'Z KTHXBAI YR model MKAY
KTHXBYE

An Ceiling Cat maded peps liek cat wat dat r lat an cul teh catz
frinished dan dey sayed, all dey pene. An Ceiling Cat maded teh
Urfs An teh way stuff dat it 2. An teh snaek sayed Ceiling Cat
be awn yu, grly. So he iz al teh moy stuff dat if u hees fiedz
En ak nok at wuz can spenteh 2 walll net din did wuz.

What Is This

A character-level transformer that learns to write LOLCODE programs and LOLCat Bible prose. The entire pipeline — model creation, training, inference — is orchestrated from LOLCODE scripts using the BRAINZ library.

The model is a 4-layer transformer with 834K parameters. It was trained on 38KB of LOLspeak text (LOLCODE programs, LOLCat Bible excerpts, and LOLspeak dialogue) for 50,000 steps using the Adam optimizer. Training takes about 90 minutes on a single CPU core.

It has no idea what it's saying, but it's saying it with confidence.

Sample Output

Prompt: An Ceiling Cat maded — it writes scripture:

An Ceiling Cat maded teh Urfs An teh waterz stuff An all teh
animulz stuff Seaz. An Ceiling Cat maded teh Urfs An teh wuz
An stuff An levz a so. An Ceiling Cat maded dem An all teh
animulz. An Ceiling Cat maded peention iz srsly.
An teh snaek sndz derey started all dat maded.
An teh snaek sayed, DO WANT an dey Fluffy wud efinty din wry.
An Ceiling Cat maded to az rd ewers write ders we

Prompt: An teh snaek sayed — it tells stories:

An teh snaek sayed Ceiling Cat be awn yu, grly. So he iz al
teh moy stuff dat if u hees fiedz En ak nok at wuz can
spenteh 2 walll net din did wuz.

Prompt: HAI 1.4 — it writes LOLCODE programs:

HAI 1.4
    HOW IZ I izprime YR i MKAY, O RLY?
             YA RLY, FOUND YR 1
        OIC
            IM IN YR loooop UPPIN YR i TIL BOTH SAEM i AN 5
        I HAS A maximum ITZ nums'Z SRS 0
    I HAS A minimum ITZ nums'Z SRS 0

Prompt: MAI KITTEH — it preaches:

MAI KITTEH ". HAH ITEH TOASTER ING?
Ceiling Cat watches baiting alked. Hey wisuz p wretetetd dem
An I ated it or ez sretly peaine. So Ceiling Cat maded teh
anywaterz dem stufeh day. An Ceiling Cat maded teh Urfs An
teh waterz stuff An I ated out

How It Works

The BRAINZ library adds native tensor operations and a transformer model to the lci LOLCODE interpreter. No external dependencies — just C and math.

Component	Details
Architecture	4-layer pre-norm transformer, single-head attention
Parameters	834K (3.2 MB on disk)
Vocab	128 (raw ASCII, character-level)
Context window	64 characters
Optimizer	Adam (beta1=0.9, beta2=0.999)
Autograd	Eager-mode tape with full backward pass
Training data	38KB of LOLspeak/LOLCODE text

Under the hood: a from-scratch autograd engine records operations on a global tape during the forward pass, then walks it backward to compute gradients. The Adam optimizer maintains per-parameter momentum estimates. All of this is in pure C — no BLAS, no GPU, no frameworks.

Usage

Prerequisites

Build lci (BRAINZ support is included by default):

git clone https://github.com/justinmeza/lci
cd lci
mkdir build && cd build
cmake .. -DCMAKE_BUILD_TYPE=Release && make

Train

Edit train.lol to adjust hyperparameters, then:

lci train.lol

STEP 0 LOSS 5.36
STEP 5000 LOSS 1.56
STEP 10000 LOSS 2.77
STEP 15000 LOSS 0.84
...
STEP 45000 LOSS 0.31
DOEN LURNIN

Chat

lci chat.lol

BRAINZ READY 4 CHATZ

U> CEILING CAT
BRAINZ> HAZ CHEZBURGER? KTHXBAI! IM IN UR CKIN TER...

U> HAI 1.4
BRAINZ> HOW IZ I izprime YR i MKAY, O RLY? YA RLY, FOUND YR 1...

U> KTHXBYE

Files

File	What it does
train.lol	Trains a new model from lolspeak.txt
chat.lol	Interactive chat REPL with a trained model
lolspeak.txt	Training corpus (LOLCODE programs + LOLCat Bible + LOLspeak)

The BRAINZ API

CAN HAS BRAINZ?                                          BTW import library

I IZ BRAINZ'Z KREEAYT YR "layers=4,hidden=128" MKAY      BTW create model
I IZ BRAINZ'Z STUDEE YR model AN YR text AN YR 0.001 MKAY BTW train step
I IZ BRAINZ'Z SPEEK YR model AN YR prompt AN YR 200 AN YR 0.8 MKAY  BTW generate
I IZ BRAINZ'Z SAVE YR model AN YR "brainz.bin" MKAY       BTW save weights
I IZ BRAINZ'Z LOAD YR "brainz.bin" MKAY                   BTW load weights
I IZ BRAINZ'Z KTHXBAI YR model MKAY                       BTW free memory

Training Corpus

The lolspeak.txt corpus is assembled from open sources:

• LOLCODE programs — FizzBuzz, Fibonacci, prime checkers, sorting algorithms, 99 bottles, Towers of Hanoi, and more from Rosetta Code and Esolangs
• LOLCat Bible — Genesis chapters in LOLspeak from the LOLCat Bible Translation Project
• LOLspeak dialogue and prose — conversations, stories, and common phrases in LOLspeak

License

WTFPL. DO WUT U WANT.