2 For the stack experiment:
6 Takes ~1h10min on a 4090.
8 For the arithmetic expressions experiments
10 # 38M parameters / 250k samples
14 # 352M parameters / 2.5M samples, reaches 99.80% after 12 epochs, the
15 learning rate schedule is obviously terrible
17 ./main.py --task=expr --nb_blocks=48 --dim_model=1024 --nb_train_samples=2500000 --result_dir=results_expr_48b_d1024_2.5M