Pi JAM

Generative language models for inference on Raspberry Pi Zero. Just Another Model(s) for the smallest Rasberry Pi!

---

Models

Variant	Type	Weights	Description
neuralpi (WIP)	Decoder transformer	~25 MB	8-layer attention model, KV cache inference
jamxe	2-layer MLP	76 KB	512→512→45, dual-source feature hash with frozen query anchor
jamkid	2-layer MLP	76 KB	Same engine as jamxe, trained on a kid-register dataset

jamxe and jamkid respond in milliseconds. neuralpi (work in progress) will be slower but architecturally more capable

---

Build

Requires gcc and make. No other dependencies for inference.

make

To cross-compile for Pi Zero (ARM1176)

CC=arm-linux-gnueabihf-gcc make

To build a single variant:

make -C neuralpi # decoder transformer
make -C jamxe    # extended MLP     (~76 KB weights)
make -C jamkid   # kid register     (~76 KB weights, same engine as jamxe)

Run

Each variant is a REPL: type a query, get a response

./jamxe/dist/jamxe jamxe/weights_xe.bin
./jamkid/dist/jamkid jamkid/weights_jamkid.bin
./neuralpi/dist/neuralpi neuralpi/weights_neuralpi.bin

When running on the Pi, the binary and weights file can sit in the same directory and the weights path can be omitted:

cd jamxe && ./dist/jamxe

Try

Variant	Try
jamxe	HOW ARE YOU, WHAT IS LOVE, 3+4, COUNT 5
jamkid	HELLO, NAME, ORCA, STORY
neuralpi	HOW ARE YOU, WHAT IS ATARI, TELL A STORY

Retrain

Training requires Python 3 and PyTorch (offline and independent from the Pi).

# MLP (jamxe / jamkid)
cd jamxe
python3 train_xe_strong.py td_xe.txt weights_xe.bin
make

# NeuralPi (transformer)
python3 train_transformer.py neuralpi/td_transformer.txt neuralpi/weights_neuralpi.bin
make -C neuralpi
# For large datasets (>10K pairs), augmentation auto-scales; override with --aug N if needed

Lineage

Lineage and naming of this repo comes from Atari JAM (Just Another Model), originally designed to run on an even tinier Atari 800.