configs/defaults.yaml

# What to inspect. CLI flags override these.

model: lerobot/smolvla_base              # HuggingFace ID or local path
dataset: lerobot/svla_so101_pickplace    # SO101 pick-and-place dataset
episode: 0
num_frames: 8
device: auto                             # auto | cpu | cuda | mps
output_dir: ./outputs
save_individual: true                   # true to also save each frame separately

# Attention analysis options
method: rollout                         # last-layer | rollout | all-layers
cross_attention: true                   # true to capture action-expert → vision cross-attention (slower)
show_heads: true                        # true to save a per-head attention grid for the first frame
raw_attention: false                    # true to skip positional baseline subtraction