appvoid/arco-mini-run-1
Updated
•
8
mini-experiment
this represent an effort to discover optimal improvements on slms using curriculum learning as n+10k samples on a 250m parameters danube model
Note final loss: 2.353
appvoid/arco-mini-run-2
Updated • 10Note final loss: 2.1243
appvoid/arco-mini-run-3
Updated • 10Note final loss: 1.7117
appvoid/arco-mini-run-4
Updated • 13Note final loss: 1.5411
appvoid/arco-mini-run-5
Updated • 12Note final loss: 1.558 (high learning rate dropped performance, working on it)
appvoid/arco-mini-run-75k
Updated • 7Note arco-mini-run-5 trained on 25k additional samples