run_large_training_lr1e4.sh

#!/usr/bin/env bash
TOKENIZERS_PARALLELISM=false python3 run_distillation_nodes.py \
  --model_name_or_path "./nb-distil-large-init-0811" \
  --teacher_model_name_or_path "NbAiLab/nb-whisper-large" \
  --train_dataset_name "NbAiLab/annotated_distil_raw_ncc_speech_v7_large" \
  --train_dataset_config_name "" \
  --train_split_name "train" \
  --eval_dataset_name "NbAiLab/annotated_distil_raw_ncc_speech_v7_large" \
  --eval_dataset_config_name "" \
  --eval_split_name "validation" \
  --eval_steps 500 \
  --save_steps 1000 \
  --warmup_steps 1000 \
  --learning_rate 0.0001 \
  --lr_scheduler_type "linear" \
  --logging_steps 200 \
  --save_total_limit 1 \
  --max_steps 50000 \
  --wer_threshold 10 \
  --per_device_train_batch_size 16\
  --per_device_eval_batch_size 16 \
  --dataloader_num_workers 16 \
  --dtype "bfloat16" \
  --output_dir "./" \
  --do_train \
  --do_eval \
  --use_scan \
  --gradient_checkpointing \
  --overwrite_output_dir \
  --predict_with_generate \
  --freeze_encoder \
  --streaming \
  --use_auth_token \
  --report_to "wandb" \
  --wandb_project "nb-distil-whisper-large-flax2" \
  --wandb_name "flax lr1e4" \
  --save_code_to_wandb \
  --save_train_state \
  --hub_model_id "NbAiLab/nb-distil-whisper-large-flax3" \
  --push_to_hub