run_large_training.sh

#!/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_compact8_large" \
  --train_dataset_config_name "no" \
  --train_split_name "train" \
  --text_column_name "text" \
  --eval_dataset_name "NbAiLab/annotated_distil_raw_ncc_speech_v7_compact8_large" \
  --eval_dataset_config_name "no" \
  --eval_split_name "validation_norwegian_fleurs" \
  --eval_text_column_name "text" \
  --eval_steps 2000 \
  --save_steps 2000 \
  --warmup_steps 1000 \
  --learning_rate 0.0001 \
  --lr_scheduler_type "linear" \
  --timestamp_probability 0.2 \
  --condition_on_prev_probability 0.2 \
  --language "no" \
  --task "transcribe" \
  --logging_steps 250 \
  --save_total_limit 1 \
  --max_steps 100000 \
  --wer_threshold 10 \
  --per_device_train_batch_size 16 \
  --per_device_eval_batch_size 16 \
  --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 \
  --report_to "wandb" \
  --gradient_accumulation_steps 2 \
  --wandb_project "nb-distil-whisper-large-test2" \
  --hub_model_id "NbAiLab/nb-distil-whisper-large-test2" \
  --push_to_hub