Paper Benchmark

← Back to benchmark overview

Fixed benchmark snapshot shared by the Hugging Face notebook, benchmark-only HTML build, and generated TeX includes. Snapshot label: paper-2026-02-11T1915Z. Cutoff: 2026-02-12T06:15:00+11:00.

Trained params is the average non-zero parameter count in the fitted model across folds. Avg active params / classification is the mean number of active parameters or scoring decisions touched while classifying one product.

6,693Filtered products

2,542Filtered tags

363Filtered taxonomies

71Prime base

View UMLLR tag-order ablation View operational UMLLR page

Benchmark dashboard chart for paper benchmark view — Small-multiples dashboard generated directly from the comparison table rows below.

Model comparison chart for paper benchmark view — Log-log scatter of trained parameters versus mean p-adic loss, overlaid with the current parsimoniousness baseline.

Active parameter comparison chart for paper benchmark view — Scatter plot of avg active params versus mean p-adic loss, excluding PCLR and PCNN. Fitted equation: mean p-adic loss = -0.0956 log10(active params) + 0.3799. Regression fitted on log10(active params): R²=0.675, p=0.023.

Active parameter versus log-loss chart for paper benchmark view — Scatter plot of avg active params versus log10(mean p-adic loss), excluding PCLR and PCNN. Fitted equation: log10(mean p-adic loss) = -0.1821 log10(active params) - 0.4503. Regression fitted on log10(active params): R²=0.794, p=0.007.

Model	Trained params	Mean p-adic loss	Exact acc.	Prefix-2 acc.	Avg active params / classification
Dummy Baseline	1.0	0.560442	5.32%	25.30%	1.00
Importance-Optimised p-adic Linear Regression	648.2	0.263237	49.82%	65.38%	1.11
Parameter-constrained Logistic Regression	11913.0	0.733943	8.35%	21.52%	560.00
Unconstrained Logistic Regression with L1	3886.0	0.085839	65.87%	83.51%	375.99
Parameter-constrained Neural Network	10999.0	0.673437	14.63%	27.33%	10149.88
Decision Tree	30947.0	0.110314	60.93%	80.12%	258.35
Level-wise Logistic Regression	123174.8	0.120450	53.81%	80.08%	199.99
Unconstrained Neural Network with L1	35209.0	0.114660	62.56%	80.90%	4766.06
Zubarev (greedy init.)	797.4	0.333238	40.49%	58.03%	1.54