dreamer_v3_200m_world_model(pretrained: bool = False, overrides: object = {})The paper's default rung with all three objectives.
Model Size
Parameters
pretrainedbool= FalseNo weights are published for this family; passing
True raises.**overridesobject= {}Optional
DreamerV3Config field overrides.Returns
DreamerV3ForWorldModelingThe rung plus the objectives.
Notes
Reference: Hafner, Pasukonis, Ba, and Norouzi, "Mastering Diverse Domains through World Models", Nature 640 (2025), 647-653 (arXiv:2301.04104).
This is the configuration that collected diamonds in Minecraft from
scratch — with, the paper is careful to note, the same hyperparameters
as everything else. What differs per benchmark is the schedule:
action repeat, environment count and replay ratio, none of which are
model fields. Pass them to lucid.utils.rollout.rollout and to
your training loop.
Examples
>>> from lucid.models import dreamer_v3_200m_world_model
>>> model = dreamer_v3_200m_world_model(action_dim=18).eval()
>>> model.config.discount, model.config.actor_entropy
(0.997, 0.0003)