tinywmcraft checkpoints
Checkpoints of tinywmcraft, a 13.9M-parameter pixel world model of Craftax-Classic trained from scratch in under 10 minutes on one A100: the benchmark's iterations 1 to 3, trained on the benchmark data (datasets/maharishiva/tinywmcraft-v3). Play the best one in your browser at maharishiva.github.io/tinywmcraft.
Each model.npz holds the tokenizer, the world model with its reward and done heads, and JSON metadata with the
training config and time; every seed's eval.py row on the public eval clips is in the repo's docs/records.jsonl. Every run was
trained cold on one Colab A100-SXM4-40GB. Score: 0.4 gFVD + 0.2 game logic + 0.2 memory (half recall, half
self-consistency) + 0.1 reward + 0.1 done, each part 0 for a trivial model and 1 for the real game, recomputed from
each row with the repo's eval.north_star; gFVD is shown times 10^4.
| folder | code | score | gFVD | logic | memory | reward AP | done AP | train s | eval s |
|---|---|---|---|---|---|---|---|---|---|
iteration1/seed{0,1,2} |
train_tiny.py at 23e6515, 178 lines |
0.287 ± 0.019 | 67.2 | 0.03 | 0.056 | 0.918 | 0.272 | 554 | 169 |
iteration2/seed{0,1,2} |
train.py at 80b1c34 |
0.463 ± 0.023 | 44.3 | 0.55 | 0.087 | 0.934 | 0.247 | 582 | 144 |
iteration3/seed{0,1,2} |
train.py and kernels.py at 8538aa5 |
0.562 ± 0.017 | 27.2 | 0.69 | 0.087 | 0.929 | 0.190 | 571 | 147 |
Means ± sample SD over the three seeds (train s and eval s: the slowest seed). Iteration 3 is about 350 lines: a one-frame tokenizer trained on clean, masked and noised frames, flow matching with a shortcut loss, an eighth of the windows starting at a real game start, Muon and two fused kernels.
curl -L --fail --create-dirs -o runs/best/model.npz https://huggingface.co/maharishiva/tinywmcraft-v3/resolve/main/iteration3/seed0/model.npz
python play_wm.py --model runs/best/model.npz
Other folders
policies/: the four RL policies and the goal walker that played the training data (collect/in the repo).web/play.wgr: iteration 3 seed 2 as one graph for the browser player (docs/release_web.sh:docs/export_web.py, thendocs/convert_web.py), fp16 weights computed in fp32, 23 MB. The site loads it at a pinned revision.scale/: a 5x larger model trained with about 10x the compute, outside the 600 s rule, from development.