box_flip_franka_mini_ood_sae

Sparse Autoencoder (SAE) Model

This model is a Sparse Autoencoder trained for interpretability analysis of robotics policies using the LeRobot framework.

Model Details

• Architecture: Multi-modal Sparse Autoencoder
• Training Dataset: jccj/mh2_step_1_flip
• Base Policy: LeRobot ACT policy
• Layer Target: model.encoder.layers.3.norm2
• Tokens: 162
• Token Dimension: 128
• Feature Dimension: 25920
• Expansion Factor: 1.25

Training Configuration

• Learning Rate: 0.0001
• Batch Size: 16
• L1 Penalty: 0.3
• Epochs: 20
• Optimizer: adam

Usage

from src.sae.trainer import load_sae_from_hub

# Load model from Hub
model = load_sae_from_hub("villekuosmanen/box_flip_franka_mini_ood_sae")

# Or load using builder
from src.sae.builder import SAEBuilder
builder = SAEBuilder(device='cuda')
model = builder.load_from_hub("villekuosmanen/box_flip_franka_mini_ood_sae")

Out-of-Distribution Detection

This SAE model can be used for OOD detection with LeRobot policies:

from src.ood import OODDetector

# Create OOD detector with Hub-loaded SAE
ood_detector = OODDetector(
    policy=your_policy,
    sae_hub_repo_id="villekuosmanen/box_flip_franka_mini_ood_sae"
)

# Fit threshold and use for detection
ood_detector.fit_ood_threshold_to_validation_dataset(validation_dataset)
is_ood, error = ood_detector.is_out_of_distribution(observation)

Files

• model.safetensors: The trained SAE model weights
• config.json: Training and model configuration
• training_state.pt: Complete training state (optimizer, scheduler, metrics)
• ood_params.json: OOD detection parameters (if fitted)

Citation

If you use this model in your research, please cite:

@misc{sae_model,
  title={Sparse Autoencoder for Mh2 Step 1 Flip},
  author={Your Name},
  year={2024},
  url={https://huggingface.co/villekuosmanen/box_flip_franka_mini_ood_sae}
}

Framework

This model was trained using the physical-ai-interpretability framework with LeRobot.