run.sh

export MODEL_NAME=distilbert-base-uncased
export MAX_SEQ_LENGTH=192

python run_glue.py \
  --model_name_or_path $MODEL_NAME \
  --dataset_name ccdv/arxiv-classification  \
  --do_train \
  --do_eval \
  --max_seq_length $MAX_SEQ_LENGTH \
  --per_device_train_batch_size 32 \
  --learning_rate 2e-5 \
  --num_train_epochs 1 \
  --max_eval_samples 500 \
  --output_dir tmp/arxiv \
  --overwrite_output_dir