s2-nano-doom

A 702 KB ternary model that plays a Doom-like shooter from 14 numbers. Same s2 architecture as s2-nano (1,532,033 parameters), specialized on a first-person shooter on a grid drawn with Freedoom art (BSD-3-Clause; no id Software assets). It reads 14 measured numbers every 4 ticks and picks one of 8 actions; a coded reflex vetoes moves into fireballs and shots with an empty weapon.

• Demo: huggingface.co/spaces/maxgfr/s2-nano-doom
• Code: github.com/maxgfr/skycmd (100 % Rust)
• Story: chapter 18, a Doom-like shooter

Results (closed loop, sky-arena, 50 episodes per game, seed 7)

game	result
Doom-like	100 % completion, kill rate 0.75 (= oracle), 0 deaths
Flappy Drone	197.6 pipes on average (2 crashes in 50)
Obstacle Dodge	72 % (oracle 78 %)

Latency: 740 µs native and 1,282 µs in WASM for one Doom question (Apple M5).

Usage

Runs in the skycmd engine only (browser, Node, Python, CLI, server):

cargo install --git https://github.com/maxgfr/skycmd sky-serve
skycmd decide --model s2-nano-doom.gguf --request req.json --pretty

Request format: install guide.

Training

From scratch in Rust (mlx-rs) on Apple Silicon (scripts/train-doom.sh): numeric masked-language pretraining, decision training with deep supervision over the exits, ternary QAT and DAgger rounds in the Doom-like game, on top of the s2 decision data (typed-decision sets, classic datasets converted into typed questions, game states labelled by exact oracles). Part of the training data was generated with GLM. Licenses: DATA_LICENSES.md.

Files

• s2-nano-doom.gguf: ternary weights, tokenizer, temperatures and exit thresholds.