code/config/train_jsb.py

# train a tiny GPT model on Bach Chorales 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 = 'jsb'
gradient_accumulation_steps = 1
batch_size = 4
block_size = 2048  # Music Transformer config

# baby GPT model :)
n_layer = 6
n_head = 8
n_embd = 256
pos_type = 'rope'
base_freq = 10000
rotate_fraction = 1.0
thetab_init = 'zero'
dropout = 0.2
bias=False

learning_rate = 6e-4 # with baby networks can afford to go a bit higher
max_iters = 3000
lr_decay_iters = 3000 # 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 = 50 # not super necessary potentially

out_dir = 'out-jsb'
eval_interval = 100 # keep frequent because we'll overfit
eval_iters = 76 // batch_size  # 76 samples in val/test 
log_interval = 10 # don't print too too often

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