Tiny memorization experiment results
Per-run learning curves (loss_curve.csv: step, epoch, lr, train_loss_bits, held_loss_bits),
final metrics (metrics.json including memorized bits, bits/parameter, L_train, L_held),
and raw state.pt checkpoints for every run. Aggregated summary.csv/summary.json
and figures (capacity_plot.png, loss_curves.png) are at the repo root.
Memorization metric (paper Sec 3.2): mem = N_data_tokens * (log2 V - L_train_bits);
bits_per_param = mem / n_params. Capacity ~ max memorization over dataset sizes.