Feature/post-processing-fixes

[fheitzer]

Hi, as I am continuing your research I had to do some changes to the post processing. I was working with the data dir that was already put on the cluster. So, I hope this is universal and not just because of the cluster data. The commits are straight forward, but ill list my changes here:

• tma column stripping to get the exact identifier had to be adjusted to eliminate all Grading prefixes and suffixes --> tma_identifiers stripped and double checked
• the free text mapping wasn't utilised at all --> utilised
• the "free text: " prefix wasn't removed --> now removed
• there still were explanations with 2 parts separated by a "\n" --> now they are separated and the first valid explanation is used
• the German / English errors mappings were not caught --> map them (German error straight to English)
• there were still double spaces in some explanations --> replaced with single spaces

The training runs for me now and the val loss decreases yippie.

[fheitzer]

Most definitely due to outdated csv