accelerate launch --mixed_precision=bf16 run_distillation.py \
  --model_name_or_path "./distil-large-v3-init" \
  --teacher_model_name_or_path "GalaktischeGurke/swhisper_large_8552" \
  --train_dataset_name "GalaktischeGurke/merge_all_OS_v3" \
  --train_dataset_config_name "default" \
  --train_split_name "train" \
  --text_column_name "sentence" \
  --eval_text_column_name "sentence" \
  --audio_column_name "audio" \
  --eval_dataset_name "GalaktischeGurke/merge_all_OS_v3" \
  --eval_dataset_config_name "default" \
  --eval_split_name "test" \
  --eval_steps 5000 \
  --save_steps 5000 \
  --warmup_steps 500 \
  --learning_rate 1e-4 \
  --lr_scheduler_type "linear" \
  --logging_steps 25 \
  --save_total_limit 5 \
  --max_steps 100000 \
  --per_device_train_batch_size 4 \
  --per_device_eval_batch_size 4 \
  --dataloader_num_workers 130 \
  --preprocessing_num_workers 130 \
  --ddp_timeout 7200 \
  --dtype "bfloat16" \
  --output_dir "./" \
  --use_pseudo_labels "false" \
  --condition_on_prev_probability "0.0" \
  --do_train \
  --do_eval \
  --gradient_checkpointing \
  --overwrite_output_dir \
  --predict_with_generate \
  --freeze_encoder \
  --streaming false \
  --push_to_hub \
  --report_to "wandb"accelerate launch --mixed_precision=bf16 run_distillation.py \
  --model_name_or_path "./distil-large-v3-init" \
  --teacher_model_name_or_path "GalaktischeGurke/swhisper_large_8552" \
  --train_dataset_name "GalaktischeGurke/merge_all_OS_v3" \
  --train_dataset_config_name "default" \
  --train_split_name "train" \
  --text_column_name "sentence" \
  --eval_text_column_name "sentence" \
  --audio_column_name "audio" \
  --eval_dataset_name "GalaktischeGurke/merge_all_OS_v3" \
  --eval_dataset_config_name "default" \
  --eval_split_name "test" \
  --eval_steps 5000 \
  --save_steps 5000 \
  --warmup_steps 500 \
  --learning_rate 1e-4 \
  --lr_scheduler_type "linear" \
  --logging_steps 25 \
  --save_total_limit 5 \
  --max_steps 100000 \
  --per_device_train_batch_size 4 \
  --per_device_eval_batch_size 4 \
  --dataloader_num_workers 130 \
  --preprocessing_num_workers 130 \
  --ddp_timeout 7200 \
  --dtype "bfloat16" \
  --output_dir "./" \
  --use_pseudo_labels "false" \
  --condition_on_prev_probability "0.0" \
  --do_train \
  --do_eval \
  --gradient_checkpointing \
  --overwrite_output_dir \
  --predict_with_generate \
  --freeze_encoder \
  --streaming false \
  --push_to_hub \
  --report_to "wandb"