Skip to content

Massive loss while training Turbo #16

Description

@Askobarus

I have been trying to finetune Turbo model for Russian language using RUSLAN dataset (precisely the first 800 files of it due to hardware restrictions (still amounting to around 1 hour of pure noisless audio)), but encountered loss around multiple 1e+12 (meaning loss is always around 10 bil to multiple trillions).
Tried with stock tokenizer.json and with tokenizer where i moved Russian characters to 'model''vocab' from 'added_tokens'. Checked metadata encoding to avoid transcription being random characters. Also tried adding max_grad_norm = 1.0 to training parameters which bore no difference.
Don't know if i am doing anything wrong or data is not suitable for training (though dataset is specifically built for TTS training purposes).

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions