This is the selected inference checkpoint from the stabilized B arm of the
independent GTN loop-attention A/B experiment. It predicts efficiency and Vout
from the graph-based topology and duty inputs used by this project.
This is a custom PyTorch checkpoint, not a Transformers checkpoint. Load it
with shared_energy.gtn_surrogate.GTNSurrogateModel from this repository.
Selected validation result
Best epoch: 662
Best validation mean RSE: 0.091561
Validation RSE, efficiency: 0.060639
Validation RSE, Vout: 0.122483
Validation MSE: 261.503524
Test evaluation was intentionally skipped, so the listed metrics are validation metrics.
The B arm resumes the common 450-epoch no-loop warmup, enables loop attention
from epoch 451, reduces the loop-phase learning rate to 3e-4, and applies global
gradient-norm clipping of 1.0. The checkpoint is restored to the best validation
epoch (662), rather than the final training epoch (700).
final_training_state.pt, periodic training_states/, and batch diagnostics are
deliberately excluded. They are resumption/debug artifacts, not inference
weights. Training history and validation predictions are uploaded only when
their explicit flags are supplied.
