BurnyCoder/eureka-digits-luna-models

eureka-digits-luna

Experimental checkpoints. Input definitions, preprocessing and usage limitations are specified by the registered protocol and reproduction guide.

Experiment	Method	Seed	Split	Accuracy	Seconds	Device
class_prior_s0	class_prior	0	test	0.102778	0.038	cpu
logistic_s0	logistic	0	test	0.955556	0.107	cpu
pilot_regularized_s0	regularized	0	validation	0.966574	7.616	cuda
regularized_s0	regularized	0	test	0.963889	0.288	cuda
regularized_s1	regularized	1	test	0.969444	6.002	cuda
regularized_s2	regularized	2	test	0.969444	2.192	cuda
sensitivity_s0	sensitivity	0	test	0.963889	2.394	cuda
sensitivity_s1	sensitivity	1	test	0.969444	2.348	cuda
sensitivity_s2	sensitivity	2	test	0.969444	2.258	cuda
shuffled_labels_s0	shuffled_labels	0	test	0.061111	2.333	cuda
shuffled_labels_s1	shuffled_labels	1	test	0.116667	2.353	cuda
shuffled_labels_s2	shuffled_labels	2	test	0.108333	2.369	cuda
unregularized_s0	unregularized	0	test	0.963889	2.235	cuda
unregularized_s1	unregularized	1	test	0.969444	2.374	cuda
unregularized_s2	unregularized	2	test	0.969444	2.323	cuda

The mean paired accuracy difference is 0.000000, with conditional 95% bootstrap interval [0.000000, 0.000000]. The registered decision is inconclusive. The sample standard deviation across seed effects is 0.000000. The interval uses 10000 paired resamples of test examples and conditions on the trained models.

• question: "Does AdamW weight decay improve held-out accuracy of a small MLP on the bundled digits data?"

• architecture: [64, 128, 10]

• optimizer: "AdamW"

• learning_rate: 0.001

• batch_size: 64

• epochs: 50

• methods: ["regularized", "unregularized", "sensitivity", "shuffled_labels", "logistic", "class_prior"]

• seeds: [0, 1, 2]

• data_boundaries: "Frozen stratified 60/20/20 train/validation/test partition. Split seed 42. Divide pixels by the documented constant 16. Pilot reports validation only. No learned preprocessing or selection uses test data. Generalization is limited to this random example split, not new writers or datasets."

• analysis_plan: "Use 10000 paired bootstrap resamples of test example indices with RNG seed 2026, preserving pairing across methods and all trained seeds. This interval conditions on the trained models and does not estimate training-population uncertainty. Separately show each seed effect and its sample standard deviation. Treat other comparisons as descriptive, with no post-hoc primary selection."

• hyperparameter_rationale: "Small prespecified network and fixed training budget provide fast feedback on an established task. The three decay values are a descriptive local sensitivity check, not a test-selected sweep. Three seeds expose limited initialization sensitivity; they do not support broad population claims."

• loshchilov-hutter-2019-adamw: Decoupled Weight Decay Regularization

• pytorch-adamw-docs: torch.optim.AdamW

• sklearn-load-digits-docs: load_digits — scikit-learn documentation

• sklearn-train-test-split-docs: train_test_split — scikit-learn documentation

• nanda-research-process-2025: How I Think About My Research Process: Explore, Understand, Distill

• lones-2024-ml-pitfalls: How to avoid machine learning pitfalls: a guide for academic researchers

• lipton-steinhardt-2018: Troubling Trends in Machine Learning Scholarship

• neurips-paper-checklist-guidelines: NeurIPS Paper Checklist Guidelines