Need set model.config.forced_decoder_ids in training mode ?

[avishai111]

Hi everyone, I have a whisper model that I want to train. The model needs to work in a specific language. I have data for training in this language.

It seems that during training the model.forced_decoder_ids field at the model.config is always set to None. At inference time, the following fields are set:
model. config.forced_decoder_ids = processor.get_decoder_prompt_ids(language="language", task="transcribe",no_timestamps=True)
I want to verify that my training is proceeding as expected, so at the validation part, I look at the preds. This is during training, so model. config.forced_decoder_ids = None. The preds seem wrong (a random text from different languages).
Setting model. config.forced_decoder_ids = processor.get_decoder_prompt_ids(language="language", task="transcribe",no_timestamps=True) during training yields logical results (the preds from the same language).
Still, on the other hand, I saw in the examples on the internet that it is not adjusted during the training but only in the inference.

I did not find when to give a value to this part during the training/validation/test. I only found information that this variable forces the model. to generate in a specific language that I choose.

I would appreciate it if someone could share some information on the subject, when to give a value in this field, and when to leave this value to None.

Thank you all,
Avishai