jinjaguo/pap-backbone

PAP backbones (50M, 200M, 800M)

Pretrained physical backbones used to build Physical Action Primitives. Part of the Physical Action Primitives collection. Code: https://github.com/Jinjaguo/physical-action-primitives.

Each size holds architecture and weights only (no optimizer state):

checkpoints/backbone/{50m,200m,800m}/checkpoint_final.pt   # keys: model_config, model_state, ...
checkpoints/backbone/{50m,200m,800m}/config.json

The repository layout mirrors $PAP_ROOT:

hf download jinjaguo/pap-backbone --local-dir $PAP_ROOT

The 800M model has 796,957,496 parameters (encoder width 2,032, 24 residual MLP blocks) and is trained on the pretraining corpus in pap-corpus.

Citation

@inproceedings{guo2027pap,
  title  = {Beyond Verbs and Trajectories: Physical Action Primitives through Local Interaction Laws},
  author = {Guo, Jinjia and Wang, Xinyu and Sun, Bowen and Zuo, Dongqi and Du, Jing},
  year   = {2027}
}