Start run on: 2026-07-17 01:05:41.989012

Feedforward network to predict alignment-augmented descendant
Ancestor sequence embedder (FULL-CONTEXT): transformer
Descendant sequence embedder (CAUSAL): transformer
Combine embeddings with: feedforward
when reporting, normalizing losses by: descendant length


2: model init

3: main training loop
New best dev loss at epoch 0: 664.8880004882812
New best dev loss at epoch 2: 635.5745239257812
New best dev loss at epoch 3: 629.762939453125
New best dev loss at epoch 4: 622.4503173828125
New best dev loss at epoch 5: 604.9095458984375
New best dev loss at epoch 7: 603.4751586914062
New best dev loss at epoch 9: 602.836669921875
New best dev loss at epoch 11: 598.1669311523438


EARLY STOPPING AT 16:

4: post-training actions
Epoch with lowest average dev set loss ("best epoch"): 11
RE-EVALUATING ALL DATA WITH BEST PARAMS

SCORING ALL TRAIN SEQS
SCORING ALL DEV SEQS
SCORING ALL HELD-OUT TEST SEQS