File size: 348 Bytes
0768ca3
de243ef
 
 
e7d5eaa
 
de243ef
1
2
3
4
5
6
7
8
Just for performing some experiments. Do not use.

Since the loss seem to start going up, I did have to restore this from 9e945cb0636bde60bec30bd7df5db30f80401cc7 (2 step 600k/200). I am then restarting with warmup decaying from 1e-4.

That did failed. Checked out c94b5bb43b05fc798f9db013d940b05b3b47cd98 instead and restarted step 3 from here.