Have you tried your dataset for other models?
#1
by
MarcGrumpyOlejak
- opened
Pardon this late question – but I have tried your dataset "jfeil/GermanDefinitionGeneration-Distillation" to set up a simple static 'base' model (only built with one dataset).
Have you tried your dataset for other models than MT5?
Did you know that it scores in a full MTEB (deu) best in comparison to Mmarco, Avemio-pairs and many others ? it also beat my mt-translated german version of gooaq (which was #1 before). It 'seems' that the mixture of your short mt-generated definition (gt) with the "Definition question" based upon the simple wiktionary content works better than even the full "wikipedia-22-12".