Datasets:

Shanmuk4622
/

msc-imagenet100

sample_idx int64 0 129k	forget_events int32 0 32	ever_correct bool 2 classes	el2n float32 0 1.41	unforgettable bool 2 classes
0	3	true	0.056958	false
1	0	true	0.05876	true
2	0	true	0.14345	true
3	2	true	0.398822	false
4	14	true	0.821589	false
5	0	true	0.008618	true
7	3	true	0.801251	false
8	1	true	0.023322	false
9	3	true	0.821975	false
10	3	true	0.742441	false
11	0	true	0.408551	true
12	1	true	0.034601	false
13	1	true	0.030947	false
14	0	true	0.054343	true
15	0	true	0.036055	true
16	0	true	0.013773	true
17	1	true	0.075532	false
18	1	true	0.036248	false
19	1	true	0.092943	false
20	9	true	0.804294	false
21	0	true	0.175005	true
22	7	true	0.970127	false
23	8	true	0.734748	false
24	2	true	0.695033	false
25	0	true	0.280808	true
26	0	true	0.009213	true
27	2	true	0.061334	false
28	5	true	0.091334	false
29	0	true	0.004474	true
31	0	true	0.255621	true
33	0	true	0.398901	true
34	8	true	1.030142	false
35	2	true	1.05684	false
36	1	true	0.180829	false
37	0	true	0.058465	true
38	3	true	0.219874	false
39	6	true	0.064931	false
40	0	true	0.038699	true
41	2	true	0.164974	false
42	9	true	0.60599	false
43	0	true	0.130776	true
44	0	true	0.607816	true
45	0	true	0.065257	true
46	1	true	0.132651	false
47	0	true	0.013817	true
48	0	true	0.041878	true
49	7	true	0.357091	false
50	0	true	0.003016	true
51	1	true	0.163401	false
52	1	true	0.114443	false
53	0	true	0.187659	true
54	0	true	0.290585	true
55	0	true	0.073282	true
56	3	true	0.111231	false
57	2	true	0.586291	false
58	10	true	1.014612	false
59	0	true	0.186039	true
60	1	true	0.417163	false
61	5	true	0.926797	false
62	0	true	0.044789	true
63	7	true	0.859753	false
64	1	true	0.221059	false
65	0	true	0.054668	true
66	1	true	0.136035	false
67	9	true	0.981393	false
68	0	true	0.003016	true
69	6	true	0.764085	false
70	5	true	0.469847	false
71	0	true	0.00152	true
72	1	true	0.350499	false
73	1	true	0.013694	false
74	4	true	0.51483	false
75	0	true	0.068742	true
76	11	true	1.098957	false
77	0	true	0.02315	true
78	0	true	0.054003	true
79	1	true	0.335027	false
80	0	true	0.010301	true
81	0	true	0.0082	true
82	1	true	0.097163	false
84	0	true	0.278212	true
85	0	true	0.275536	true
86	0	true	0.260086	true
87	1	true	0.053422	false
88	7	true	1.025082	false
90	5	true	0.059266	false
91	3	true	0.523907	false
92	3	true	0.127887	false
93	0	true	0.013547	true
94	0	true	0.180024	true
95	0	true	0.327168	true
96	7	true	0.161902	false
97	0	true	0.035175	true
98	9	true	0.617801	false
99	0	true	0.017601	true
100	0	true	0.006388	true
103	2	true	0.515771	false
104	5	true	0.33528	false
105	0	true	0.005846	true
106	0	true	0.006414	true

End of preview. Expand in Data Studio

MSC · ImageNet-100

Per-sample Minimum Sufficient Compute for ImageNet-100 at 224px: the cost-normalised compute each sample needs before its decision has settled. Companion to the CIFAR-100 study at Shanmuk4622/msc-cifar100.

Seed reliability (rho_seed, tau=0.1)

architecture	family	seeds	rho_seed	Jaccard@10	top-1
`resnet50`	resnet	2	0.8220	0.6208	0.8237
`vit_small_p16`	vit	2	0.6492	0.2777	0.6079

rho_seed is the Spearman correlation between the MSC of two seeds of the same architecture. It is the noise ceiling: no transfer claim can exceed it, and every disattenuated number in the paper divides by it.

Layout

runs/{run_id}/
  config.yaml  config_hash.txt  STATUS.json  summary.json
  metrics/     epochs.csv  final.csv  confusion_matrix.csv  per_class.csv
               exit_metrics.csv
  telemetry/   energy_samples.csv  system_samples.csv  step_traces.jsonl
  per_sample/  test.parquet  train_holdout.parquet  train_dynamics.parquet
               meta.json
  checkpoints/ ckpt_last.pt  ckpt_best.pt
  env/         environment.json
  exit_heads.pt

budgets/{arch}.json   registry/   analysis/   tables/   paper/figures/

per_sample/test.parquet is the scientific artifact; everything in the paper is computed from it. sample_idx is the global pack index, not a position within a split, so tables from different runs join directly.

Run identifiers

{phase}-{arch}-{dataset}-{method}-s{seed} — e.g. p0-resnet50-imagenet100-base-s1. Phase p0 is the pilot, p3 MSC-KD.

Caveats

The pilot is 2 architectures x 2 seeds. rho_seed has no error bar.
No architecture appears in both the CIFAR and ImageNet studies, so cross-study magnitudes confound architecture, resolution and dataset. The CNN > ViT ordering replicates; the magnitudes do not transfer.
MSC-KD is a negative result: at matched FLOPs it is below confidence thresholding, and the oracle ceiling shows there was no headroom.

Full detail: docs/imagenet100/24_IN100_STATUS.md in the code repository.

Downloads last month: 221