lower lr regarding explode gradient. also new seed number

run.sh

@@ -12,14 +12,15 @@ python3 run_mlm_flax.py \
     --preprocessing_num_workers="64" \
     --per_device_train_batch_size="32" \
     --per_device_eval_batch_size="32" \
-    --learning_rate="3e-4" \
-    --warmup_steps="1000" \
+    --learning_rate="3e-5" \
+    --warmup_steps="2000" \
     --overwrite_output_dir \
-    --num_train_epochs="10" \
+    --seed="19" \
+    --num_train_epochs="8" \
     --adam_beta1="0.9" \
     --adam_beta2="0.98" \
     --logging_steps="500" \
     --save_steps="10000" \
-    --eval_steps="10000" \
+    --eval_steps="20000" \
     --dtype="bfloat16" \
     --push_to_hub