Wine Dataset

Parameter efficiency across Linear Softmax and Boolear architectures on a classical, linearly separable benchmark.

Published: August 2026

Setup

Dataset

Wine (UCI)

Features / Classes

13 features, 3 classes

Samples

~178

All three models reach 100% test accuracy. The question is just how many parameters it takes to get there.

Model	Params	Train	Val	Test
Linear Softmax	42	97.6%	100%	100%
Leaf V6	27	97.6%	100%	100%
Leaf V8 (most compact)	16	94.4%	96.3%	100%

Leaf V8 achieves identical test accuracy with 16 parameters — 2.6× fewer than Linear (42), 1.7× fewer than Leaf V6 (27). This is the most compact model.
Leaf V8 appears to underfit on train/val (94.4% train vs 97.6% for the others, 96.3% val vs 100%), yet generalizes perfectly on the test set. This is a meaningful signal: the model hasn't memorized the training distribution but has learned something structurally correct about the class boundaries.
Leaf V6 matches Linear exactly on all three splits despite having 37% fewer parameters — they find the same decision boundary through different means.