code/config/train_maestro.py

# train a tiny GPT model on the MAESTRO dataset

# we expect to overfit on this small dataset, so only save when val improves
always_save_checkpoint = False

wandb_log = False # override via command line if you like
wandb_project = 'complex-rope'
wandb_run_name = 'tiny-gpt'

dataset = 'maestro'
augment = True
gradient_accumulation_steps = 1
batch_size = 16
block_size = 2048

# baby GPT model :)
n_layer = 8
n_head = 8
n_embd = 256
pos_type = 'rope'
base_freq = 10000
rotate_fraction = 1.0
thetab_init = 'zero'
dropout = 0.1
bias=False

learning_rate = 6e-4 # with baby networks can afford to go a bit higher
max_iters = 60000
lr_decay_iters = 60000 # make equal to max_iters usually
min_lr = 6e-5 # learning_rate / 10 usually
beta2 = 0.99 # make a bit bigger because number of tokens per iter is small
weight_decay = 1e-2
warmup_iters = 500 # not super necessary potentially

out_dir = 'out-maestro'
eval_interval = 1000 # keep frequent because we'll overfit
eval_iters = 760 // batch_size  # split sizes -> non-augmented: 12619/698/764; augmented: 98k/698/764
log_interval = 10 # don't print too too often

complex_flash = False # use custom complex_flash_attention
# on macbook also add
# device = 'cpu'  # run on cpu only
# compile = False # do not torch compile the model