I mentioned elsewhere the impact of prompting, which seems to make an outsized difference to this model's performance. I tried NER and POS tagging (with somewha

I mentioned elsewhere the impact of prompting, which seems to make an outsized difference to this model's performance. I tried NER and POS tagging (with somewhat disappointing results). One thing that worked strikingly well was translation on non-Indo-European languages. Like I had success with Thai and Bahasa Indonesian -> English...