gtn-base-v3

GTN Base V3 Cleaned Surrogate

This repository contains the gtn-base-v3 cleaned GTN surrogate checkpoint used in the shared_energy experiments.

Files

• gtn_surrogate_model.pt: PyTorch checkpoint.
• manifest.json: training/evaluation metadata.
• training_history.jsonl: per-epoch training and validation history.
• validation_predictions.jsonl: validation split predictions.
• test_predictions.jsonl: test split predictions.

Model

• Canonical class: shared_energy.gtn_surrogate.GTNSurrogateModel
• Role: GTN surrogate for cleaned power-topology simulation data.
• Targets: efficiency, vout
• Loss: relative squared error (rse)

Training Data

• Dataset type: JSONL
• Sample count: 57,024
• Split mode: random
• Train rows: 45,619
• Validation rows: 5,702
• Test rows: 5,702

Training Summary

• Completed epochs: 545
• Stopped early: true
• Best epoch: 465
• Best validation RSE mean: 0.0119547601
• Final validation RSE:
  • efficiency: 0.0030197722
  • vout: 0.0208897479
• Final test RSE:
  • efficiency: 0.0031667461
  • vout: 0.0198842511

Notes

This is a research artifact for the shared_energy power-topology workflow. Use it together with the repository code that defines shared_energy.gtn_surrogate.GTNSurrogateModel.