Toturial of finetune?
Hello osman. Thank you for sharing the model. I take your suggestion and converted the text from uas to uls. I use the peft_bnb_whisper_large_v2_training for the finetune process. The training procese goes well, hovever the loss stuck at 0.7. I tried with small lr and warm_up step but does not help. The evaluation setp in the trainer keep giving error so I removed the evaluation during training. I tested the model after training and found the model only give a single "é
" with blank. I was wondering do you have any process for the text or audio except resample? Could you give a simple toturial? Thanks!
Im using common_voice_16 ug, here is the converted tsv.
Hi, I have not faced such a problem. Have you used Uzbek tokeniser after the training?
Yes, I'm using Uzbek tokenizer and precessor. I found that your training goes down smooth to a quite small value(0.0073 at 4000 step). But mine stuck at 0.7. Dont know what's the problem.