Before each run, I wrote a prediction. That small ritual mattered more than I expected: it turned tuning from a search for pleasing numbers into a record of which intuition survived.

The baseline stayed fixed at seed 42, batch 32, context 128, four layers, four heads, width 128, and 2,000 steps. Then I changed one main variable at a time.

Plate V.1 — context length

Prediction. More context should improve validation loss, while costing more memory and time.

ContextTrain tokensBest valWallPeak memory
644.10M1.83327.8s127 MB
1288.19M1.78227.2s205 MB
25616.4M1.74929.9s497 MB

The direction matched the prediction, but the explanation did not stay clean. At the same number of steps, doubling T also doubles the number of training tokens. Context 256 did not merely see farther; it saw four times as many token positions as context 64.

Correction. This experiment says the larger-context configuration won under a same-step budget. It does not isolate context length as the cause.

Plate V.2 — depth

Prediction. More layers should improve validation loss and cost substantially more time.

LayersParametersBest valWall
20.42M1.91618.3s
40.82M1.78227.2s
61.21M1.73438.2s

The token budget stayed at 8.19M, so this attribution is cleaner. Depth helped and the wall time rose with it.

Plate V.3 — width

Prediction. More width should help, but parameter count should grow quickly.

WidthParametersBest valWallPeak memory
640.21M2.03627.9s177 MB
1280.82M1.78227.2s205 MB
2563.21M1.62428.4s420 MB

On this A100 and this small workload, width produced the largest validation improvement while wall time stayed nearly flat. The cost appeared in parameter count and memory instead.

This is not a perfectly pure width knob either. I kept n_head=4, so increasing n_embd also increased each head’s dimension from 16 to 32 to 64.

That is a local result, not a scaling law. The honest sentence is: width won among these settings, on this corpus, under this budget.

Prediction / result / correction

Prediction. More context, depth, width, and training should all help in increasingly obvious ways.

Result. All four could improve the score, but each attached a different bill: extra tokens, time, parameters, memory, or overfitting.

Correction. “More” is not an explanation. A result earns meaning only after the budget and confound are printed beside it.

The cleanest number in an experiment is often the caveat that prevents it from becoming a slogan.

Unpaid debt

The teaching model is now measurable. The next temptation is to equate a larger training system with a different intelligence. Plate VI: The Extra Code Is Not Another Brain separates the next-token spine from the factory around it.