# 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: cuda # auto | cpu | cuda | mps task: null # override task/language instruction (default: from dataset) output_dir: ./outputs/mthirumalai/ save_individual: true # true to also save each frame separately image_map: null # dataset→policy image key mapping, e.g. "front=camera2,side=camera3" # 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: true # true to skip positional baseline subtraction attn_threshold: 0.5 # percentile (0-1) below which attention values are zeroed # Gradient-based attribution gradient: both # null | saliency | gradcam | both gradient_device: cuda # null (same as device) | cpu | cuda | mps gradient_seed: 42 # fixed noise seed for reproducible gradient attribution smooth_grad: 20 # SmoothGrad samples (1 = vanilla saliency, >1 = averaged) smooth_grad_sigma: 0.15 # Gaussian noise std for SmoothGrad # Extended attribution features gradcam_connector: true gradcam_vlm_layers: "4,8,12,16" per_step_cross_attention: true vision_vs_state: true per_action_dim: true language_diff: true # Data export (for web viewer) export_data: true # save structured .npz + JSON alongside PNGs run_name: null # run folder name (default: timestamped run_YYYY-MM-DD_HH-MM-SS) # Model internals report internals_only: false with_internals: true internals_frames: 5 entropy_warn: 0.8 entropy_critical: 0.95 entropy_low: 0.1 redundancy_warn: 0.7 redundancy_critical: 0.9