Objective: To characterize a silent training-collapse failure mode in parameter-efficient fine-tuning (PEFT) of large vision transformers (ViTs) for small-sample industrial computed tomography (CT) defect recognition, and to identify the evaluation protocol that masks it.
Methods: A DINOv2 ViT-L/14 (304M) was fine-tuned with LoRA on a 673-image lung CT dataset; a single-variable ablation isolated the root cause; the signature was reproduced on LIDC-IDRI (6,691 slices); and a matched 2×2 learning-rate control tested an apparent MAE domain-adaptation degradation.
Results: The model silently collapsed to a non-learning majority-class predictor (accuracy 32.4%, cross-entropy 1.370, just below the uniform floor ln(4) ≈ 1.3863). The collapse was masked by best-accuracy tracking: a best-epoch snapshot of 51.11% hid a final model at the 32.4% prior. Lowering the learning rate from 10⁻³ to 2×10⁻⁴ alone prevented collapse (frozen-norm ViT-L reached 94.20%), and the apparent +11.15% LayerNorm advantage reduced to an insignificant −1.19%. On LIDC-IDRI, a ViT-B/14 collapsed on 2/3 folds (62.24%, F1 = 0, AUC = 0.5), and fixing the protocol recovered 96.85%. A matched learning-rate control showed a reported −28.7 percentage-point MAE degradation to be a learning-rate artifact, with MAE-adapted LoRA reaching 92.6% at the stabilized rate.
Limitations: The ln(N) floor is specific to classification; only LoRA was tested; datasets number hundreds to thousands of samples; and the industrial validation relies on synthetic simulated data.
Conclusions: The headline normalization effects are artifacts of training instability, not genuine LayerNorm effects. Practitioners should report final-epoch metrics alongside best-epoch metrics, treat loss convergence to ln(N) as a collapse diagnostic, and use a scale-aware learning rate.
