Brief
Train a CIFAR10-style subset classifier with frozen-backbone + unfreeze ablation. Submit accuracy + confusion matrix + ablation table.
Dataset
CIFAR10-style subset; train-only augmentation statistics enforced. Trap: test-leak via augmentation is graded.
Starter guardrails (88)
- Slots: frozen baseline, unfreeze run, ablation table, confusion matrices.
- Test transform deterministic; normalization stats from train only.
Deterministic checks
Accuracy >= 0.70 on the hidden test slice; confusion matrices present for both runs; augmentation-stats provenance stated.
100-pt rubric
- Frozen baseline quality: 20
- Unfreeze ablation rigor: 25
- Accuracy bar met: 15
- Confusion-matrix analysis (per-class deltas): 25
- Leak-free augmentation provenance: 15