These two steps were run several times, because some of the jobs exceeded walltime, and I was trying to fix the time using the `task.attempt` flag, but that didn't work. 

Also, when I first ran, I forgot about all the SGD parameters and ran all three optimisers, which is not great. Again, why there are so many trace files. 

So there are quite a few trace files. Other than the last one (nf-trace.txt), the 3rd last one is the last attempt with SGD, the merged report with raw_full flag is the one that grabs all jobs, inc the SGD ones. This is just for record. All of these jobs used 6205 SU, where as the ones that did ran used 3139 SU. 

The work directory and the rest of the trace files are probably no longer needed, but just kept for now. 



TODO: in the finetune directory of all the models and results, the SGD folders are all there. Need to delete them all. Probably better to move all the Adam ones and then delete what's left.