starrydark/fec-final-weights

Fusion Equilibrium Challenge 2026: Final reproduction weights (team AmadeusBio)

Trained weights for the team's three Final submissions. Code, configs and the training and inference entrypoints: https://github.com/starryark/fec26-final-package. The package pins this repository at revision dfa0ee10f2064c1c34f8523e522507bb50620532 and checks every file it downloads against its SHA-256 (weights_manifest.json in the package). Use that revision; later commits on main change only this README and the copy of the manifest.

Final submission	package commit (git tag, branch)	config
FW-899AF2ED	476bbf08 (FW-899AF2ED, v11)	final_MTfull_mmk
FW-E0D18468	476bbf08 (FW-E0D18468, v11)	final_MTfull
FW-7700D079	3a3663fd (FW-7700D079, v12.1)	final_A22li_dec_mmk_nrsl

Contents at the pinned revision

path	what	used by
d3d/d3d_li2ftA22mt_* (22)	DIII-D members trained with auxiliary shape and current-density heads, then li(2)-fine-tuned	final_MTfull, final_MTfull_mmk
d3d/d3d_li2ftA22_* (22)	DIII-D members li(2)-fine-tuned without auxiliary heads	final_A22li*, incl. final_A22li_dec_mmk_nrsl
d3d/d3d_b60ax*, d3d/d3d_bxax* (28)	older no-Thomson ensemble: the fallback for frames without Thomson data	every config
d3d/d3d_bxw3e_* (28)	width-3072 base members of earlier candidates	final_A24, final_B, final_C
offdec/ (3)	off-span decoders	final_A22li_dec*, final_MTfull*, final_MTli
mast/n7h_f0.55/	amortized MAST network (model.pt SHA-256 bfbb346c…)	every config with network MAST maps
cache/ (2)	passive-model caches for the MAST vessel currents	MAST stages
synthetic/gen_v7h_f0.55/ (34)	synthetic MAST equilibria generated from public MAST inputs only (no labels, no EFIT outputs): the training set of the Development-phase network. A full rebuild regenerates the set by default, which is how the shipped network was trained (package README, "Expected rebuild variance")	train.py --synthetic only

Inference (bash reproduce.sh <config>) downloads every file listed in the package's manifest (108 files, 96 GB).