genie_coinrun(pretrained: bool = False, overrides: object = {})Construct the CoinRun case study — Genie at single-accelerator scale.
Model Size
Parameters
pretrainedbool= FalseNo weights were released;
True raises.**overridesobject= {}Optional
GenieConfig field overrides.Returns
GenieModelThe three networks, untrained.
Notes
Reference: Bruce et al., arXiv:2402.15391, Appendix F — Table 15 for the tokenizer, Table 16 for the latent action model (six actions), Table 17 for the dynamics model (sampling at temperature 1). The appendix trains the tokenizer on batches of 48 sixteen-frame clips and the other two, together, on batches of 36.
Three values it does not give are filled in: 64x64 frames, a latent
action patch of 16, and 8 dynamics heads where Table 17 prints
num_layers twice. The appendix reports no parameter count; this
build has 188M.
Examples
>>> from lucid.models import genie_coinrun
>>> model = genie_coinrun()
>>> config = model.config
>>> config.frame_shape, config.num_latent_actions, config.temperature
((64, 64), 6, 1.0)
>>> config.token_grid
(16, 16)