CaliBench: Model Logits and Features from ResNet and DenseNet

CaliBench contains extracted logits and features from ResNet-50, ResNet-110, and DenseNet-121 models trained on CIFAR10, CIFAR100, and ImageNet. Developed by zhurong2333 and updated in March 2026, the collection facilitates research into model calibration and focal loss performance.

Use Cases

• Evaluating post-hoc calibration techniques using logits from ResNet-110
• Comparing feature representations between DenseNet-121 and ResNet-50 on ImageNet
• Analyzing the impact of focal loss on logit distributions in CIFAR100

Strengths

• Includes outputs from three distinct architectures: ResNet-50, ResNet-110, and DenseNet-121
• Covers three standard benchmarks: CIFAR10, CIFAR100, and ImageNet
• Provides logits specifically from focal loss calibration experiments

Limitations

• Contains derived model outputs rather than raw source images
• Documentation lacks specific column headers and data schema

Provenance

Source

zhurong2333 via Hugging Face

Collection Method

Extracted: Features and logits were extracted from pretrained PyTorch models during inference.

Freshness

Last updated March 2026.