#!/usr/bin/env bash
accelerate launch run_distillation.py \
  --model_name_or_path "./nb-distil-large-init" \
  --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 0.0003 \
  --lr_scheduler_type "constant_with_warmup" \
  --timestamp_probability 0.2 \
  --condition_on_prev_probability 0.2 \
  --language "no" \
  --task "transcribe" \
  --logging_steps 200 \
  --save_total_limit 1 \
  --max_steps 50000 \
  --wer_threshold 20 \
  --per_device_train_batch_size 32 \
  --per_device_eval_batch_size 32 \
  --dataloader_num_workers 8 \
  --preprocessing_num_workers 8 \
  --ddp_timeout 7200 \
  --dtype "bfloat16" \
  --attn_implementation "sdpa" \
  --output_dir "./" \
  --do_train \
  --do_eval \
  --gradient_checkpointing \
  --overwrite_output_dir \
  --predict_with_generate \
  --freeze_encoder \
  --freeze_embed_positions \
  --streaming True \
  --wandb_project "nb-distil-whisper-large-pytorch"
  --wandb_name "pytorch2_lr3e-4"
  --push_to_hub