Regarding NER finetuning

[malduwais]

Dear Authors,

I am trying to re-produce the NER results using WikiANN. However, I didn't get the same results even I am using the same hyperparameter for XLM-R and mBERT models.

Is there something I need to set other than the hyperparameter?

I am using Trainer method as follow:

args = TrainingArguments(
model_path,
overwrite_output_dir=True,
num_train_epochs=10,
load_best_model_at_end=True,
evaluation_strategy="epoch",
save_strategy="epoch",
report_to="wandb",
seed=seed,
data_seed=seed,
per_device_train_batch_size=32,
per_device_eval_batch_size=32,
learning_rate=2e-5,
weight_decay=0.01,
warmup_ratio=0.1,
)

Thank you,