lvesucces/Dual_Tower_Selector_Weights

Query-Conditioned Dual-Tower Safety Rule Selector Weights

This model repository contains the trained weights and hyperparameter specifications for the Query-Conditioned Dual-Tower Safety Rule Selector (Experiment 1 of the ICRA 2027 paper framework).

---

1. Model Overview

• Architecture: Query-Conditioned Dual Tower MLP
• Function: Given a task-scene query $Q_t$, dynamically projects candidate safety rules and computes normalized hyperplane distance $d_{t,i} > 0$ to select applicable rules from the constraint space.
• Training Metrics (dual_tower_hyperplane_bce_all_metrics.json):
  • F1 Score: 0.9948
  • Precision: 0.9896
  • Recall: 1.0000 (100% ground-truth safety rule recall)
  • Accuracy: 0.9995

---

2. Included Weights Files

• dual_tower_hyperplane_bce_all.pth: Primary Query-Conditioned Hyperplane Dual-Tower model weights (39.9 MB)
• dual_tower_bce_all.pth: Standard BCE Dual-Tower model weights (39.9 MB)
• dual_tower_svm_margin_all.pth: SVM-Margin Dual-Tower model weights (39.9 MB)
• whole_mlp_bce_all.pth: Whole MLP Baseline weights (1.1 MB)
• dual_tower_hyperplane_bce_all_metrics.json: Training history, evaluation metrics, and network architecture configuration.
• dual_tower_weights_bundle.zip: Zipped archive of all model weight files and metrics.