johnypuzzles
Landing page
A puzzle you play in the terminal. You connect over SSH, and a small language model, the
Bordadeira (the embroiderer), stitches poems about the Portuguese flag for you. One of her poems
is hiding something. [Not Finished btw, still spits a lot of gibberish]
The answers are not in this repository. They are read from environment variables, so you can
run your own box with your own.
How to play
- Connect to the box with
ssh. It runs on demand, so the address is shared when it's open.
- Press enter at the coat of arms and read the riddle (in Portuguese).
- Type anything and the Bordadeira embroiders a poem.
- Take her home if you like: the box serves her model over
scp.
- Submit with
code <answer>. Three tries.
How it's made
| Part | What |
|---|
| The Bordadeira | Qwen2.5-0.5B-Instruct + a LoRA adapter (not finished :P), fine-tuned on Camões (Os Lusíadas and the lyric poems, on public domain https://huggingface.co/datasets/vreabernardo/camoniana) plus AI generated poems about the PT flag. Small enough for a laptop CPU. |
| Training | TU Delft's DelftBlue cluster, Slurm jobs on an A100, deterministic seeds and an evaluation gate before shipping (just for fun tbh, no need for A100s). |
| The box | An asyncssh server: no password, a welcome screen that redraws to any terminal size, per-IP limits, the adapter served read-only over SCP. |
| Hosting | A small Azure VM with a systemd service and on/off scripts, so it only costs money while people play. |
| Visuals | The coat of arms drawn from text glyphs, as ANSI in the terminal and as a canvas on the landing page. |
Run your own box
Python 3.10+.
git clone https://github.com/joaofilipe5/johnypuzzles && cd johnypuzzles
python3 -m venv .venv && .venv/bin/pip install -r requirements.txt
export JP_CODE=YOURWORD JP_ANSWER=YOURANSWER JP_SEED=42
.venv/bin/python -m server.ssh_server --port 2222
ssh -p 2222 localhost # in another terminal
The server creates its own host key on first run. Without a trained adapter the box still runs
and the Bordadeira "rests". To train one see docs/DELFTBLUE.md; to host it
see infra/DEPLOY.md.
Repository
server/ the SSH box players connect to
model/ bordado: dataset, LoRA training and evaluation
vexnet: a small character-level model trained on flag descriptions
render/ the coat of arms as ANSI text and as a web canvas
co/ an earlier optimisation-based version of the puzzle (max-cut)
hpc/ Slurm job scripts for DelftBlue
infra/ create, deploy, start and stop the Azure VM
docs/ the landing page and the DelftBlue guide
Trained weights, the player bundle and the training data are not included.
