Training changes weights; inference uses them Diagram

Educational pretraining, not a foundation-model run

Train a tiny randomly initialized model only to see batches, shifted targets, loss, gradients and optimizer updates.
The useful capstone reuses a small public pretrained checkpoint. Do not pretrain a billion-parameter model on this Mac.

The simplest post-training experiment

Supervised examples teach short domain labels. LoRA updates a small set of adapter parameters while the original base weights remain frozen.
Optional continued pretraining, distillation, preference optimization and reinforcement learning are distinct choices, not compulsory steps.

A falsifiable result

Compare the unchanged baseline and the adapter on the same untouched examples, prompt, output budget and hardware.
Measure macro-F1, errors, invalid outputs, time to first token, total latency and memory. A lower training loss alone is insufficient.

System 1-like is an operating goal

A small specialized model produces a bounded decision without a generated reasoning essay. This is not a new transformer component.
Quantization reduces representation precision; KV caching reuses inference state. Neither is equivalent to fine-tuning.