RiverRider/srt-cxr14-frozen-probe

Frozen general-purpose features beat a fine-tuned baseline on ChestX-ray14 A linear probe on frozen google/gemma-4-31B-it hidden states, on all 112,120 images of NIH ChestX-ray14, using the official test_list.txt. No fine-tuning, no radiology pretraining, no augmentation. One Linear(d, 14) under BCE, which is fourteen logistic regressions. ChestX-ray14, official split mean AUROC method Wang et al. 2017 (dataset authors) 0.7451 ResNet-50, fine-tuned end to end this… See the full description on the dataset page: https://huggingface.co/datasets/RiverRider/srt-cxr14-frozen-probe.