все вопросы только сюда @flexyzitarian (никаких подтасовок в коде только честность)

Решение задачи: Noisy Parity с помощью архитектуры EliaOmega V3 (Argent Fold)

=== Datasets === === Model === Parameters: 256,823 res_scale=0.5774 x0_decay=[1.0, 0.8633841276168823, 0.7454321384429932] === Training (500 epochs) === Epoch 001/500 | loss=0.8816 | train=50.0% | test=50.9% | lr=2.50e-05 | scale=1024 | t=6s Epoch 002/500 | loss=0.6994 | train=49.9% | test=50.6% | lr=5.00e-05 | scale=4096 | t=13s Epoch 003/500 | loss=0.6961 | train=49.7% | test=49.4% | lr=7.50e-05 | scale=32768 | t=19s Epoch 004/500 | loss=0.6948 | train=50.0% | test=50.3% | lr=1.00e-04 | scale=131072 | t=26s Epoch 005/500 | loss=0.6944 | train=50.0% | test=50.9% | lr=1.25e-04 | scale=262144 | t=32s Epoch 006/500 | loss=0.6940 | train=50.3% | test=52.4% | lr=1.50e-04 | scale=262144 | t=38s Epoch 007/500 | loss=0.6938 | train=50.3% | test=52.5% | lr=1.75e-04 | scale=524288 | t=44s Epoch 008/500 | loss=0.6631 | train=56.3% | test=91.0% | lr=2.00e-04 | scale=131072 | t=50s Epoch 009/500 | loss=0.5329 | train=77.0% | test=99.9% | lr=2.25e-04 | scale=131072 | t=57s

GROKKING on epoch 9!

Done. Best: 99.9% | NaN batches: 0 | Time: 56.5s