Work out dataset size requirement instantly with clear inputs, formula shown and shareable results.
Classical learning theory ties sample complexity to model capacity: with roughly ten samples per free parameter, a model has enough evidence to fit each weight without memorising. The rule is crude for deep networks, where regularisation, augmentation and pretraining shift the requirement by orders of magnitude, but it remains a useful floor for tabular models and a warning sign when a network is far larger than its data.
Rule of ten
samples required = parameters x samples per parameter; per class = samples / classes; add a holdout fraction on top
No. Pretrained weights already encode most of the structure, so instruction fine-tuning can succeed with a few thousand high-quality examples despite billions of parameters.
Shrink the model, add regularisation, use augmentation or transfer learning, and prefer simpler feature-based models over deep architectures.