Bad performance when fine-tuning Vicuna on Arabic data

Nobody has claimed this yet.

Hello,

Given that Vicuna actually responds to Arabic prompts in Arabic (although lacking), I still judged it to be suitable enough for fine-tuning on Arabic data (more specifically, instructional data and closed-book QA data). However, after fine-tuning the 7b variant, the performance is really bad as it's not even able to construct a sentence that's fully correct when prompted.
I expected a mediocre performance after seeing that the loss was in the range of 4-6, but what could be the reason behind this? Note that I modified the Llama tokenizer so that it supports Arabic vocabulary by integrating an Arabic sentencepiece model.

Any help is most appreciated! thanks

No repository file, test, or entry point is named. Start by reproducing fine-tuning of the Vicuna 7B variant on the described Arabic instructional and closed-book QA data, then inspect the modified Llama tokenizer and reported loss range; done requires identifying and documenting the cause of the poor Arabic output.

Written by the indexing model from the issue text.