Hi @AudranBert We need few more information to debug this. Could you share your previous tarred dataset config and current lhotse dataloading config and also information on model you are trying to train. Along with sample manifest line CC: @pzelasko

Hi, thanks for the response!
As I said, I try to finetune this hybrid french model. My config files and a sample of my last bucket (the data are a bit too heavy to be sent but i can try to send the audios of the 3 segments) for tarred dataset are attached in the configs.zip file
configs.zip

I compared both configs and they look fine. No major red flags. Few points to debug:

1. I see you have override_config_path: model_v2.yaml not sure what you have in them. Better to consolidate into one single yaml file
2. You can reduce num_workers: 2
3. Remove quadratic_duration: 30 default should be enough
4. Could you try re-estimating buckets with the script: https://github.com/NVIDIA/NeMo/blob/main/scripts/speech_recognition/estimate_duration_bins.py
5. Minor: change compute_eval_loss to False (this doesn;t affect but to save memory)

If those doesn;t work I would recommend you to turn off bucketing and replicate simple configuration first and then take it from there.

Ok thanks, I will try to test when I can. For 1. it is just the same config file as the one of the model I try to finetune. It is useless now but at the beginning I used to change a few values. I will remove it, thanks