any training details?

#1
by seastar105 - opened

it's good to see new hubert model trained on korea, could you inform training details(e.g. hyperparams, dataset) about this model?

Sorry for the late confirmation
Currently, the model card has been updated, so please check it.

I will close the issue

hyunwoo3235 changed discussion status to closed

thanks for reply.

could you explain more about which dataset is used for training kmeans model for second training?
like in hubert paper, authors used subset of librispeech

seastar105 changed discussion status to open

Similarly, we randomly sampled 100 hours from the combination of datasets we used and used them to train the kmeans model.

@hyunwoo3235 thanks for sharing, what was final loss and top-k accuracy of training? for both base and large model

The final losses for the base and large models were 1.995 and 2.25, respectively.
unfortunately, the model was not tested as it was trained as a validation for another project. therefore, there are no other performance metrics.

seastar105 changed discussion status to closed

Sign up or log in to comment