README.md

metadata

license: cc-by-4.0
tags:
  - pruning
  - compression
  - benchmarks
  - sparsification
  - sal
size_categories:
  - n<1K

sal-torch Benchmarks

Raw results from sal-torch experiments: SAL-trained vs standard-trained models under compression.

Files

• multiseed_results.json — GPT-2 Medium / SST-2, 5 seeds, full 872-example validation split
• robustness_scale_results.json — robustness suite across model scales
• jepa_sal_multiseed_summary.json — I-JEPA ViT-H/14 / ImageNet-100, 3 seeds
• sal_qat_prototype_seed{42,123,456}.json — SAL+QAT on ViT-B/16 / ImageNet-100

GPT-2 Medium / SST-2 (5 seeds, mean accuracy)

Setting	Standard	SAL	Δ	SAL wins
dense	90.05%	90.62%	+0.57pp	4/5
INT8	88.88%	89.61%	+0.73pp	4/5
INT4 (NF4)	90.02%	90.23%	+0.21pp	3/5
prune 33% heads	85.71%	88.17%	+2.45pp	5/5
prune 50% heads	80.87%	82.94%	+2.06pp	4/5
prune 33% + INT8	82.84%	84.72%	+1.88pp	4/5
prune 33% + INT4	85.67%	87.25%	+1.58pp	5/5

Other results

Model	Result	Seeds
I-JEPA ViT-H/14 (631M)	SAL wins 35/36 comparisons at 33/50% head pruning (+0.4 to +1.9pp); no effect at 70%	3
ViT-B/16 SAL+QAT	+22pp mean at 50% pruning + INT4	3

Links

• 📦 sal-torch on PyPI — pip install sal-torch
• 🔧 GitHub
• 🌐 cognitive-engineering.dev
• 🛡️ sal-certify
• 📄 PPAM 2026, Springer LNCS (Paper CP031)

Data: CC BY 4.0. sal-torch: BSL 1.1. Built by Cognitive Engineering (Switzerland).