# base training dataset for training AF3 design models (atom14 variants):
# protein subsampling only.

defaults:
  - design_base
  - val/sm_binder_hbonds@val.sm_binder_hbonds
  - val/sm_binder_hbonds_short@val.sm_binder_hbonds_short
  - _self_


# Datasets:
train:
  pdb:
    probability: 0.9
    sub_datasets:
      interface:
        dataset:
          transform:
            crop_spatial_probability: 1.0

  # Should we do monomer for hbond? probably not, maybe small sampling so model learns something?
  monomer_distillation:
    probability: 0.1