Saving train state of step 1000

{checkpoint-1000 → checkpoint-2000}/flax_model.msgpack

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:a04f4f83603f30a2ef2cda5b555cf138fde78de8b57c0d754bf8de44e7a2e76c
+oid sha256:d2c82c4afd0d59fcd3b10af4269ed282aad4d6b1a59ebd6cf9f8178d4838ffa4
 size 1512831199

{checkpoint-1000 → checkpoint-2000}/train_state.msgpack

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:7c7ad83a7e9cb5ca3e7b8aba9b6aaebf06c17d0d0ccea63ca51127348a443bff
-size 7564063736
+oid sha256:4432f109f2f5c00cd4efb39749e2e876e584b9473b53018baa10b23491631ac0
+size 1216344064

run_large_training_lr9e4.sh

@@ -0,0 +1,41 @@
+#!/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_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 100 \
+  --save_steps 1000 \
+  --warmup_steps 1000 \
+  --learning_rate 0.0009 \
+  --lr_scheduler_type "linear" \
+  --logging_steps 100 \
+  --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 lr1e9 wer10" \
+  --save_code_to_wandb \
+  --save_train_state \
+  --hub_model_id "NbAiLab/nb-distil-whisper-large-flax4" \
+  --push_to_hub
+