Share the training codes

#1
by keithhon - opened

Hi, would it be possible to share your training codes for this model? I wonder how the resulted cer is such low..

keithhon changed discussion status to closed

I literally followed the tutorial from the blog

https://huggingface.co/blog/fine-tune-whisper

keithhon changed discussion status to open

did you multiply the wer by 100?

keithhon changed discussion status to closed

It was a very preliminary study on the prospect of using Whisper to transcript YouTube videos for language model usage. Didn't modify anything.

The results wasn't up to expectations so I didn't continue pursuing it.

I see. Becoz the tutorial’s WER is 3x%, and yours is 0.x%… maybe the fine tuned model over fitted too much

I found the bash scripts I used. The model was trained for 50,000 steps on 2 GPUs, which is 20 times higher than the tutorial. I think the model is overfitted....

I see… thanks

Btw can you contact me on LinkedIn (link on my HF profile) to discuss some other projects related to Cantonese

Sure, invitation sent

Sign up or log in to comment