Saving weights and logs of step 10000

README.md

@@ -1 +1,7 @@
 Just for performing some experiments. Do not use.
+
+This needed to be restarted at 100k. I am getting memory errors at the end of the epoch. Not really sure why. 
+
+Step 2 is therefore on train_2__4. Static learning rate for a while. The first 100k ended at 0.59. This is decent so early. No point in running more epochs here though. Changing the corpus and continue training.
+
+

events.out.tfevents.1641426799.t1v-n-e1a08808-w-0.871181.0.v2

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f1fe9d08288d5eebd9d406223ced869b7f987018053a53aaf0ad8e5da3cf4c11
+size 1470136

flax_model.msgpack

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:5b892d1acff31af7d49006a03d609301575fed7b9753d8361d9404bea6459e84
+oid sha256:6a85b53846a6926416b78941df5f65df77a3dfe44ecedecc4352158a10c7daec
 size 498796983

run_step2.sh

@@ -0,0 +1,29 @@
+./run_mlm_flax.py \
+    --output_dir="./" \
+    --model_type="roberta" \
+    --model_name_or_path="./" \
+    --config_name="./" \
+    --tokenizer_name="./" \
+    --train_file /mnt/disks/flaxdisk/corpus/train_2_4.json \
+    --validation_file /mnt/disks/flaxdisk/corpus/validation.json \
+    --cache_dir="/mnt/disks/flaxdisk/cache/" \
+    --max_seq_length="128" \
+    --weight_decay="0.01" \
+    --per_device_train_batch_size="192" \
+    --per_device_eval_batch_size="192" \
+    --learning_rate="4e-4" \
+    --warmup_steps="0" \
+    --overwrite_output_dir \
+    --num_train_epochs="1000" \
+    --adam_beta1="0.9" \
+    --adam_beta2="0.98" \
+    --adam_epsilon="1e-6" \
+    --logging_steps="10000" \
+    --save_steps="10000" \
+    --eval_steps="10000" \
+    --preprocessing_num_workers="64" \
+    --auth_token="True" \
+    --static_learning_rate="True" \
+    --dtype="bfloat16" \
+    --push_to_hub
+