sh-islam/ciln-bench-voters

CILN-Bench voters

These are the trained classifiers, called voters, that labeled the CILN-Bench datasets. Each voter was trained on the clean-label-train split of its dataset. It never saw the rows it later labeled.

Path	Voter	Dataset
image/resnet20_cifar10_best.pt	ResNet-20	CIFAR-10
image/wrn28_10_cifar10_best.pt	WRN-28-10	CIFAR-10
image/deit3_small_cifar10_best.pt	DeiT3-Small, fine-tuned from ImageNet-21k weights	CIFAR-10
not a file	CLIP ViT-B/32, zero-shot, prompt "a photo of a {class}"	CIFAR-10
image/lenet5_mnist_best.pt	LeNet-5	MNIST
image/mlp_mnist_best.pt	MLP	MNIST
image/resnet20_mnist_best.pt	ResNet-20	MNIST
image/deit3_small_mnist_best.pt	DeiT3-Small	MNIST
tabular/xgboost_adult_dummyna.json	XGBoost	Adult
tabular/catboost_adult.cbm	CatBoost	Adult
tabular/mlp_adult_best.pt	RTDL-style MLP	Adult
tabular/ft_transformer_adult_best.pt	FT-Transformer	Adult
not a file	TabPFN, used as is	Adult
text/fasttext/model.bin	fastText	AG-News
text/distilbert/	DistilBERT-cased, Hugging Face save_pretrained folder	AG-News
text/roberta-base/	RoBERTa-base, Hugging Face save_pretrained folder	AG-News
not a file	all-MiniLM-L6-v2, zero-shot with one prompt per class	AG-News

The train_summary.json files next to the checkpoints give the final validation and test accuracy of each run.

MIT license.