kkaulen/ctrain_pareto_fronts

Pareto-front models for Rethinking Evaluation Paradigms

This repository contains 145 checkpoints on the final, completely verified per-method Pareto fronts reported in Rethinking Evaluation Paradigms in IBP-based Certified Training (ICML 2026). Each dataset/network setting uses its canonical paper verification budget; alternate-budget comparison fronts are not mixed into the model set. model_manifest.json records each model's experiment metadata, reported accuracies, byte size, and SHA-256 digest.

The checkpoints are CTRAIN bounded-model state dictionaries. Use the utility in ADA-research/CTRAIN to reconstruct the correct architecture and wrapper:

from papers.rethinking_evaluation_paradigms.model_hub import (
    list_models,
    load_model,
)

models = list_models(
    repo_id="kkaulen/ctrain_pareto_fronts",
    dataset="cifar10",
    architecture="cnn7",
    method="sabr",
    epsilon=2 / 255,
)
model = load_model(
    repo_id="kkaulen/ctrain_pareto_fronts",
    config_hash=models[0]["config_hash"],
    device="cuda",
)
logits = model(images)

Install CTRAIN and its Git-hosted bound-propagation dependencies before loading a model:

pip install CTRAIN
ctrain-install-git-deps
pip install huggingface-hub

Checkpoint organization

checkpoints/<dataset>/<architecture>/<epsilon>/<method>/<config_hash>.pt

The weights are distributed under the CTRAIN repository's MIT license. Dataset terms remain those of CIFAR-10, MNIST, and Tiny ImageNet.

Citation

@inproceedings{KauEtAl26,
  title = {Rethinking Evaluation Paradigms in IBP-based Certified Training},
  author = {Kaulen, Konstantin and Shavit, Hadar and Hoos, Holger H},
  booktitle = {Proceedings of the 43rd International Conference on Machine
               Learning (ICML 2026)},
  year = {2026}
}