This was the first output that felt exciting (excerpted from one longer sample):
ROMEO:
Fit that that please the me distred he besty:
[...]
First ORIOLANENCE:
He wilt we as gace thee? worther thee more mine
[...]
QUEEN SLIZABETH:
No, I shall I can god and his doubtager
It had character names, colons, line breaks, and the visual rhythm of a play. It also did not know English.
That combination made the sample useful. The model learned format before spelling, and a plausible silhouette before reliable language.
Plate IV.1 — the complete desk card
Three tests, three different claims
1. Can it memorize one batch?
I trained a smaller diagnostic model on the same fixed batch for 201 updates. Loss fell from 4.2077 to 0.0104.
That proves the forward pass, loss, backward pass, and optimizer can cooperate. It does not prove generalization. A system that cannot memorize a tiny batch is broken; a system that can is merely eligible for the next test.
2. Does validation loss keep improving?
The baseline moved steadily:
| Step | Train | Validation |
|---|---|---|
| 0 | 4.2003 | 4.1982 |
| 400 | 2.3069 | 2.3175 |
| 1,000 | 1.9347 | 2.0130 |
| 1,600 | 1.6982 | 1.8422 |
| 2,000 | 1.6089 | 1.7817 |
The train/validation gap widened, but validation was still at its minimum when the budget ended. This run had not yet shown the clear rebound I saw in a different classroom configuration. Those scores cannot be compared as if the models, context, and training budget were identical.
3. Can the checkpoint reproduce the model?
I loaded a fresh model and optimizer from the saved checkpoint, fed the same prompt, and used torch.testing.assert_close on the logits. It passed. The file stored step 2000 and 52 optimizer-state entries.
That verifies loading. It is not yet seamless resume: an exact training continuation would also need RNG state and the data-loader position.
Prediction / result / correction
Prediction. If loss falls and the sample resembles Shakespeare, the model has learned language.
Result. The diagnostics proved progressively more specific things: the graph is live, held-out prediction improved, state can be restored, and local form emerged before spelling.
Correction. No single green check owns the word “learned.” Evidence has to keep its scope.
QUEEN SLIZABETHis not a failure to hide. It is the cleanest specimen of what the model learned first.
Unpaid debt
The baseline gives me an instrument. Now I can turn one knob at a time and see what moves. Plate V: Width Paid, Context Lied, Last Step Lost is the experiment that changed my intuition most.