maharishiva/tinywmcraft-v3

tinywmcraft data

Data of the tinywmcraft benchmark: train a world model of Craftax-Classic in under 10 minutes on one A100, beat the best score, then play it in your browser.

file	contents
train.npz	500,000 transitions: 250 streams of 2000 steps, 63x63 RGB frames. The only data a submission may train on. sha256 c900266cd31bf5c2a3577b87ac721a89f67b0c25052cc8a744c75c64731aa340
eval.npz	the public eval clips eval.py scores: play, game logic, and memory walks with their recall and self-consistency checks (make_eval.py in the repo). sha256 f81b0e69012071aead3ec65459abd49c2b8a2c764be75d8b99a3fc2b8035b773

Submissions are judged on a private test set built the same way from other worlds; it is not published.

train.npz

array	shape
frames	(2001, 250, 63, 63, 3) uint8
actions, rewards	(2000, 250)	actions[t, e] and rewards[t, e] take frames[t, e] to frames[t+1, e]
dones	(2000, 250) bool	deaths only (802)
truncated	(2000, 250) bool	the end of a piece that is not a death (508)
starts	(2000, 250) bool	the first frame of a piece that begins at a real game start (1020 pieces)
meta		the collection recipe, the pools and the event counts

A stream is a row of pieces. A piece ends at a death or on a quiet transition (no reward, no inventory or vitals change, not inside a return trip) and the next piece starts there, often in the middle of an episode. An episode or piece ends at dones OR truncated; train a done head on dones only, and sample game starts from starts.

Collection (collect/ in the repo): four RL policies play, a pixel PPO agent fine-tuned for diamonds drives most episodes; a behaviour schedule breaks their habits (tool order, crafts right after a furnace, failed and repeated attempts) and adds deaths of six kinds. About every third step of free play the agent makes a return trip: a peek of 2-3 tiles and back, or a loop of 4 to 63 tiles, walked by a planner or a learned goal walker, so places leave the view and come back. Pieces are picked so that events, not frames, are balanced: the cells that come back into view count as events by how long they were gone, and so does work that succeeds (chop, mine, drink, eat, place, craft), mostly by day. 316 diamonds (281 first in an episode), 802 deaths, every one of the 22 achievements.

7.1% of the cells entering the view at its edge come back within 32 frames, presses that do nothing take 4% of the frames, the agent moves in 56% of them, and the iron sword comes first in 67% of the pieces that make both iron tools.

eval.npz

Every clip starts with 8 real frames and the action into each; the model then generates one frame per action.

group	clips
play_*	2048	64 actions of the data driver's play; 512 hold a diamond, an iron tool, a plant eaten or a death early on
logic_*	1752	8 scripted actions from a real state next to something to use, 7 families
memory_*, long_*	512, 256	walks out and back, 48 and 144 actions: *_checks are cells seen in the context that come back into view, *_self cells the model draws itself that leave and come back
tile_tex, tile_night, hud_tpl		the game's textures the probes read frames with