CIFAR-10H Disagreement Predictor

Upload an image and the model will predict a 10-way human annotator distribution rather than a single hard class. The output approximates q(y|x), the predicted distribution of labels that a pool of human annotators might assign.

Loaded checkpoint: /app/checkpoints/demo_model.pt

• loss: kl
• backbone_init: cifar10_pretrained
• head: mlp
• device: cpu