test

run_large_nodes_training.sh

@@ -0,0 +1,38 @@
+#!/usr/bin/env bash
+TOKENIZERS_PARALLELISM=false python3 run_distillation_nodes.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" \
+  --eval_dataset_name "NbAiLab/annotated_distil_raw_ncc_speech_v7_compact8_large" \
+  --eval_dataset_config_name "no" \
+  --eval_split_name "validation_norwegian_fleurs" \
+  --eval_steps 10 \
+  --save_steps 1000 \
+  --warmup_steps 100 \
+  --learning_rate 0.0001 \
+  --lr_scheduler_type "linear" \
+  --logging_steps 25 \
+  --save_total_limit 1 \
+  --max_steps 10000 \
+  --wer_threshold 10 \
+  --per_device_train_batch_size 16\
+  --per_device_eval_batch_size 16 \
+  --dataloader_num_workers 32 \
+  --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-test3" \
+  --hub_model_id "NbAiLab/nb-distil-whisper-large-flax1-no" \
+  --push_to_hub
+