← Tutorials14 / 17
Build a tiny LLM on your CPU
What does the 3-character model produce?Tap to reveal
Loss drops to 0.487, and the samples are real words and sentence shapes, learned from just 632 characters in under a second on a CPU.
The loss wobbles between epochs (0.509 → 0.539 → 0.487) because each update uses one example at a time.
3) mlp, context 3 epoch 10 loss 0.585
3) mlp, context 3 epoch 40 loss 0.487
sample: the sun. / a bird sat still. /
a small. / a dog. / the mat. /
the seed on the park.