Configuration and CLI

Command-Line Reference

Run python transformer_from_scratch.py --help for the authoritative parser help.

Argument	Default	Effect
--smoke-test	none	Offline synthetic CPU test; mutually exclusive mode
--train	none	Small Yelp experiment; mutually exclusive mode
--full	none	Large Yelp preset; mutually exclusive mode
--train-size	1000	Requested train count for --train
--val-size	200	Requested validation count for --train
--test-size	200	Requested test count for --train
--batch-size	8	DataLoader batch size
--max-len	128	Fixed token sequence length
--epochs	2	Training epochs
--n-layers	3	Custom Transformer encoder blocks
--n-heads	6	Attention heads; must divide model width
--dropout	0.05	Dropout probability
--lr	0.0002	AdamW learning rate
--seed	42	Python/PyTorch manual seed
--sdpa	off	Use PyTorch SDPA attention path
--visualize	off	Optional torchview/Graphviz PNG
--output	project-local transformer_checkpoints/	Checkpoint / visualization directory

Shape constraints

d_model is inferred from the frozen embedding matrix. d_model % n_heads == 0 must hold (384 % 6 == 0 real-data defaults; 48 % 4 == 0 smoke test).