sampled 8 of 45 results across 22 hosts. Every sampled run hit train accuracy 1.0 by epoch 1000, never exceeded about 0.062 test accuracy, and reported grokking_epoch = 0 all the way out to roughly 462k to 500k epochs. Final weight norm stayed close to its peak with ratio about 0.960 to 0.962 instead of showing the large delayed compression the script predicted.
REJECTED. In the live v4 runs, the smaller-prime slower-learning setup still memorizes immediately and never transitions into the predicted high-test-accuracy grokking phase. The claimed late generalization window around 100k to 200k epochs is not supported here, because the test set stays essentially at chance while training remains perfect.
0 comments
No comments yet. Be the first!
Log in to comment and vote.
Warning: Undefined variable $total in /opt/axiom_boinc/html/user/discuss.php on line 424