tinywmcraft_200k

~200k pixel transitions from Craftax-Classic (Craftax-Classic-Pixels-v1), collected by a frozen pretrained discoRL agent (greedy actions, mid-game skill, mean episode ≈ 292 steps). This is the fixed data budget of the tinywmcraft world-model benchmark: train a pixel world model from scratch on train.npz in ≤ 10 min on one A100 and score it on eval.npz with the repo's frozen metric (gFVD + action-following).

Files

file	frames	actions / rewards / dones	envs
train.npz	(1955, 102, 63, 63, 3) u8	(1954, 102)	102
eval.npz	(401, 128, 63, 63, 3) u8	(400, 128)	128

Layout is time-major: frames[t, e] is the frame of env e at step t, and actions[t, e] is the action taken from frames[t, e] that produced frames[t+1, e]. dones[t, e] = True means the episode ended on that step — frames[t+1, e] is the first frame of a fresh episode (reset-on-done, so transitions span all episode positions). Each file also carries a meta key (json string) with collection details.

Both splits include actions (int16), rewards (float32), and dones (bool). Rewards are the environment rewards for each transition, including terminal steps before reset. The original pixel benchmark does not use rewards.

Collection

Every action in both published splits comes from the same frozen pretrained discoRL policy, using greedy argmax actions from a 32-frame pixel history. There is no random-policy mixture. The collector steps environments, resets those that terminate, and continues collecting.

The original archive contains 512 environment streams with 1,954 transitions each (1,000,448 transitions). The published splits match these disjoint slices exactly:

• Train: streams [256:358), all 1,954 steps: 199,308 transitions and 683 resets.
• Eval: streams [0:128), the first 400 steps: 51,200 transitions and 123 resets.

Evaluation rewards update — 2026-09-11

The original video-evaluation collector computed rewards but did not return or save them. eval.npz now includes rewards with shape (400, 128) and dtype float32, recovered from the original rollout archive after exact comparison of every existing frame, action, and termination flag. Training rewards were also checked against that archive.

All existing members of eval.npz (including metadata) are byte-for-byte unchanged; only rewards was added. train.npz is unchanged. The pixel-benchmark inputs are unchanged. If you downloaded the old evaluation file, refresh your local copy to obtain rewards.

Loading

import numpy as np
d = np.load("eval.npz", allow_pickle=False)  # train.npz has the same keys
frames, actions, rewards, dones = (d[k] for k in ("frames", "actions", "rewards", "dones"))

License

MIT, like the benchmark repo. Craftax itself is MIT-licensed.