kshitijthakkar/nirnaya-0.4b-0.2a-decision · 0.2B active · one forward pass per decision

Nirnaya Arcade

Ten camera, drawing and text games decided by a 0.4B typed-decision model. Your browser turns frames into facts, Nirnaya reads the facts and answers a typed question, and an independent referee says whether it was right.

Decision Lab

Decision Lab

320 held-out rounds with rule or source references. Commit, reveal Nirnaya's distribution, flip one fact.

Run it locally

On your own machine every decision skips the network. Python 3.11 or newer; a CUDA GPU is used when PyTorch sees one, otherwise the CPU.

1 · get the code

git clone https://huggingface.co/spaces/kshitijthakkar/nirnaya-arcade-v2
cd nirnaya-arcade
python -m venv .venv

2 · install and start

.venv/bin/pip install -r requirements.txt      # Windows: .venv\Scripts\pip
.venv/bin/python -m server.main                 # Windows: .venv\Scripts\python

Open http://127.0.0.1:7860. The first start downloads the model (about 0.9 GB) and warms it up.