tiny-doom-defender

A 1.1M-parameter ModernBERT encoder that plays VizDoom's Defend the Center in real time on a CPU from pixels alone.

📝 Read the full story/see the model play live · 💻 GitHub repo

Inspired by VAGOsolutions/SauerkrautLM-Doom-MultiVec-1.3M.

The model underwent SFT from a programmatic oracle and has been refined with PPO.

Checkpoints in this repo

Checkpoint	Path	Training	Mean kills	σ	Size
tiny-doom-defender (default)	root	SFT + PPO	23.12	2.81	4.4 MB
tiny-doom-defender-sft	sft/	SFT only	22.19	3.76	4.4 MB
tiny-doom-defender-int8	int8/	SFT + PPO, int8	23.13	2.78	1.1 MB

SauerkrautLM-Doom-MultiVec-1.3M (the inspiration) kills 20.38 enemies, with a standard deviation of 5.35.

Evaluation on 1000 episodes, seeds 10000–10999.

Usage

Install the package, download the repo, and let it play:

pip install git+https://github.com/anakin87/tiny-doom-defender
hf download anakin87/tiny-doom-defender --local-dir tiny-doom-defender

# watch it play in a live DOOM window
play-doom --ckpt tiny-doom-defender

# score it on the held-out test seeds
eval-model --ckpt tiny-doom-defender --episodes 100

Point --ckpt at a subfolder to use the other checkpoints:

eval-model --ckpt tiny-doom-defender/sft  --episodes 100
eval-model --ckpt tiny-doom-defender/int8 --episodes 100

Architecture

Parameters	1,096,134
Observation	3 stacked 160×100 RGB frames (t, t−1, t−2) + the 2 previous actions
Stem	two stride-2 3×3 convs, 32 channels → a 40×25 = 1000-token grid
Encoder	4-layer ModernBERT, hidden size 128, 4 heads, fed via inputs_embeds
Head	learned attention pooling → turn (3) / shoot (2) / value heads
Action space	MultiDiscrete([3, 2]) — turn left/none/right × shoot/no-shoot