Fine-tuned model forgets prior knowledge
#81
by ukemamaster - opened
Hi,
I am following your fine-tuning guidelines to fine-tune the model (microsoft/Phi-4-multimodal-instruct) for ASR task, for custom words in Spanish language. The issue is that the model forgets prior knowledge and invents words and phrases in random cases.
Did anyone notice this? Does your fine-tuned model forget prior knowledge?
Thanks
Can you please share your finetuning script?
I am following this script: https://colab.research.google.com/drive/1JAQdpX3BtIgDmTLlnHgstKfGw7HjSfej?usp=sharing