Correct held-out perplexity to full-split test 1.4369 / 4.21 (was 60-batch log numbers)
#7
by Compactbot - opened
No description provided.
Merging: corrects held-out perplexity from the 60-batch training-log numbers (val 1.9046 / test 1.9473) to the honest full-split measurement (loss 1.4369 / ppl 4.21 on 49,674 test tokens). Text-only change.
Compactbot changed pull request status to merged