SEAL circuit paper: selected Task-Net and Loss-Net checkpoints
Original, byte-preserved checkpoints for the completed September 11, 2026 single-duty comparison. This release contains four Task-Net models (T0–T3) and four paired Loss-Net variants (R0–R3), totaling nine .pt files. R0 has separate efficiency and gamma files; R1–R3 contain both heads in each checkpoint.
Models and evaluation
| Task | Paired Loss-Net | Selected Task epoch | Direct NGSpice development Utility |
|---|---|---|---|
| T0 | R0 | 670 | 0.660341 |
| T1 | R1 | 363 | 0.665198 |
| T2 | R2 | 417 | 0.661540 |
| T3 | R3 | 459 | 0.665879 |
All Task models use seed 0, the same S02 model and AdamW starting state, learning rate 0.0003, batch size 128, and 700 additional epochs. EarlyFusionMatchedSupportTaskGTN predicts one duty in [0.1, 0.9] from topology and target. Training uses duty MSE + 0.02 × (1 − frozen Loss-Net predicted Utility). Loss-Net contains EfficiencyGTN and GammaGTN physical heads.
- R0: original heads, stored efficiency epoch 690 and gamma epoch 590.
- R1: replay-only, selected update 10,000.
- R2: random-local, selected update 4,096.
- R3: Task-local, selected update 10,000.
- R1–R3 were each trained through 20,000 updates per head. This release contains the actually adopted checkpoints, not the final update checkpoints.
Development evaluation uses 99 typed topologies and 327 (topology, target) contexts over 10 targets, with equal weighting across targets. Checkpoints were selected by their own frozen Loss-Net's predicted development Utility; direct NGSpice did not select checkpoints. This development population is not an independent test set.
T3 has the highest seed-0 development mean, but superiority was not confirmed on the additional 63-topology / 120-context confirmation population: T3 minus T2 = +0.003909, adjusted 98.3333% confidence interval [-0.012061, +0.020145]. T1 was a development-only added control. These results do not establish Dynamic superiority or unseen-connectivity generalization.
Files and loading
manifest.json records pairings, exact paths, SHA-256 hashes, checkpoint keys, source-relative provenance, and W&B run links. SHA256SUMS can verify the nine original files. Source code: seal-circuit-paper, revision 2591e6762f5f5044b15d1534801d90f02720c784.
These are project-specific PyTorch checkpoints, not Transformers from_pretrained models. Reconstruct the matching architecture using the linked code and original package names. Task checkpoints store model, completed_epochs, and selection_score; they do not contain optimizer/RNG state for resuming training. R0 stores each head under model; R1–R3 store the two heads under models and also retain optimizer/RNG/provenance state. No checkpoint metadata was rewritten. Historical source paths embedded in original checkpoints are provenance, not portable file dependencies.
Example download (loading and model reconstruction require the project environment):
from huggingface_hub import hf_hub_download
path = hf_hub_download(
repo_id="DanielJeongsooLee/seal-circuit-paper-checkpoints",
filename="task-net/T3/selected.pt",
)
This release excludes datasets, Task recovery checkpoints, dual-duty models, and the later incomplete alternating3 campaign. No new license grant is asserted by this model card.