Queimo/rem3di-combisolv-ablation-no-decorrelation

Three-epoch component ablation

Variant: no_decorrelation.

Each run starts from the exact evaluated PCQMep2→COSMOep10 checkpoint. Data, geometry, order/shuffle seeds, initial task heads, optimizers and first-three-epoch learning rates match the full control. Training stops after three epochs. BigSolDB uses the identical 9686 test rows, 75731 training contexts, all1281 columns and Kumo Large four estimators, seed42.

The denoising objective corrupts atomic embeddings, not coordinates. For no_denoising, the decoder receives no gradients or updates; its loss and complete uncertainty term (including log variance regularizer) are excluded. A detached diagnostic forward preserves the original dropout/noise RNG sequence. For no_decorrelation, all three UW task terms remain, and the covariance coefficient alone is zero.

Variant	Epoch	MAE	RMSE	R²
full_model	1	0.569691	0.744798	0.572648
full_model	2	0.571346	0.748295	0.568627
full_model	3	0.576577	0.749793	0.566898
no_decorrelation	1	0.574835	0.751849	0.564518
no_decorrelation	2	0.570582	0.747797	0.569200
no_decorrelation	3	0.575272	0.752397	0.563884

One matched training seed; checkpoint3 is the predeclared final comparison, with all epoch results retained. No test-based hyperparameter changes.

W&B