Start here: the first table row is one complete three-file CE+SD demonstration — before logits, after logits and correct labels. The following rows cover every published arm, including scope-labelled (best) results and the best CE-only control. Click an arm for its explanation, recipe and full epoch/step table.

All accuracies below are ImageNet-1K validation Top-1. Peak is not the same as final. Both companion archives, shared labels, reports and exclusions are in this table. The embedding archive now contains 16 arrays: seven encoder states with support/validation coverage plus the two retained development caches; the October raw-head TTT rows link their unchanged matching embeddings and logits.

Experiment / arm	Before: full logits	After: full logits and every retained epoch/step	Checkpoint + optimizer / global parent	Validation Top-1: peak versus final
START HERE — one CE+SD run, exactly three files	logits_before.npy — fitted epoch 0	logits_after.npy — CE+SD epoch 10; correct_labels.npy	No model loading needed; compare the three arrays directly	80.586% → 82.634% (+2.048 pp)
EMBEDDINGS — Original pretrained/reference backbone	Training support features — 1,231,154 rows	Validation features — 50,000 rows; IDs, labels and explanation	Associated global checkpoint	768-dimensional FP16 features, not class logits; Loading example
EMBEDDINGS — CE-only — epoch 20	Training support features — 1,231,154 rows	Validation features — 50,000 rows; IDs, labels and explanation	Associated global checkpoint	768-dimensional FP16 features, not class logits; Loading example
EMBEDDINGS — CE+SD — selected epoch 10	Training support features — 1,231,154 rows	Validation features — 50,000 rows; IDs, labels and explanation	Associated global checkpoint	768-dimensional FP16 features, not class logits; Loading example
CE-only — epoch 20 + MUON, LR 0.005	Step 0	Step 20; embeddings	Starting global checkpoint only; no adapted query states	Final 82.366% (+0.282 pp); scalar trajectory peak 82.448% @ step 10
CE-only — epoch 20 + MUON, LR 0.02	Step 0	Step 20; embeddings	Starting global checkpoint only; no adapted query states	Final 82.096% (+0.012 pp); scalar trajectory peak 82.432% @ step 3
CE-only — epoch 20 + SGD, LR 0.001	Step 0	Step 20; embeddings	Starting global checkpoint only; no adapted query states	Final 82.224% (+0.140 pp); scalar trajectory peak 82.224% @ step 20
CE-only — epoch 20 + SGD, LR 0.01	Step 0	Step 20; embeddings	Starting global checkpoint only; no adapted query states	Final 82.098% (+0.014 pp); scalar trajectory peak 82.194% @ step 1
CE+SD — selected epoch 10 + MUON, LR 0.005	Step 0	Step 20; embeddings	Starting global checkpoint only; no adapted query states	Final 83.234% (+0.602 pp) (best) — this eight-configuration final-step grid; scalar trajectory peak 83.266% @ step 13
CE+SD — selected epoch 10 + MUON, LR 0.02	Step 0	Step 20; embeddings	Starting global checkpoint only; no adapted query states	Final 83.154% (+0.522 pp); scalar trajectory peak 83.222% @ step 3
CE+SD — selected epoch 10 + SGD, LR 0.001	Step 0	Step 20; embeddings	Starting global checkpoint only; no adapted query states	Final 83.066% (+0.434 pp); scalar trajectory peak 83.100% @ step 18
CE+SD — selected epoch 10 + SGD, LR 0.01	Step 0	Step 20; embeddings	Starting global checkpoint only; no adapted query states	Final 82.948% (+0.316 pp); scalar trajectory peak 83.022% @ step 3
Original-teacher CE+SD — main run Fast5 → LR `1e-4` (best)	E0: shared before CE / CE+SD	All E1–10, E15–20; E11–14 logits absent; E10 peak; E20 final	All epochs; E10 peak; E20 final	82.634% @ epoch 10 (best) — global training; final 82.624% @ epoch 20
CE-only — main supervised control Fast5 → LR `1e-4`	E0: shared before CE / CE+SD	All E1–20; E14 peak; E20 final	All epochs; E14 peak; E20 final	82.170% @ epoch 14 (best) — CE controls; final 82.086% @ epoch 20
Shared fitted reference — before every original CE and CE+SD arm	Same starting point, not a separate CE-only baseline	Epoch 0 logits; All reference aliases	Fitted reference model (head-fit state is not a global SGD continuation)	80.586%, direct-image evaluation
Original-teacher CE+SD — lower LR floor Fast5 → LR `1e-6`	E0: shared before CE / CE+SD	All E1–15; E14 peak; E15 final	All epochs; E14 peak; E15 final	82.294% @ epoch 14; final 82.284% @ epoch 15
CE-only — lower LR floor Fast5 → LR `1e-6`	E0: shared before CE / CE+SD	All E1–15; E14 peak; E15 final	All epochs; E14 peak; E15 final	81.622% @ epoch 14; final 81.620% @ epoch 15
Original-teacher CE+SD — higher LR floor Fast5 → LR `1e-3`	E0: shared before CE / CE+SD	All E1–15; E6 peak; E15 final	All epochs; E6 peak; E15 final	81.676% @ epoch 6; final 80.006% @ epoch 15
CE-only — higher LR floor Fast5 → LR `1e-3`	E0: shared before CE / CE+SD	All E1–15; E6 peak; E15 final	All epochs; E6 peak; E15 final	80.154% @ epoch 6; final 78.164% @ epoch 15
CE10 → CE, repeat the schedule Repeat Fast5 → LR `1e-4`	CE epoch 10	All E11–20; E20 peak	All epochs; E20 peak	80.868% @ epoch 20; final 80.868% @ epoch 20
CE10 → CE+SD, refresh teacher to CE10 Repeat Fast5 → LR `1e-4`	CE epoch 10	All E11–20; E20 peak	All epochs; E20 peak	81.766% @ epoch 20; final 81.766% @ epoch 20
CE+SD10 → CE+SD, refresh teacher to CE+SD10 Repeat Fast5 → LR `1e-4`	CE+SD epoch 10	All E11–20; E19 peak; E20 final	All epochs; E19 peak; E20 final	81.976% @ epoch 19; final 81.950% @ epoch 20
Historical CE — slower cosine to 1e-6	E0: shared before CE / CE+SD	All E1–10; E10 peak	All epochs; E10 peak	81.208% @ epoch 10; final 81.208% @ epoch 10
Historical CE — slower cosine to 1e-4	E0: shared before CE / CE+SD	All E1–10; E10 peak	All epochs; E10 peak	80.840% @ epoch 10; final 80.840% @ epoch 10
Historical CE — constant LR 0.005	E0: shared before CE / CE+SD	All E1–10; E1 peak; E10 final	All epochs; E1 peak; E10 final	74.990% @ epoch 1; final 68.086% @ epoch 10
Head-only TTT from original reference — SGD 0.01	Step 0: matched cached baseline	All Steps 0–20; Step20 peak	Starting global checkpoint only; no adapted query states	81.432% @ step 20 (best) — original-reference SGD grid; final 81.432% @ step 20
Head-only TTT from original reference — SGD 0.001	Step 0: matched cached baseline	All Steps 0–20; Step19 peak; Step20 final	Starting global checkpoint only; no adapted query states	81.228% @ step 19; final 81.228% @ step 20
Head-only TTT from original reference — SGD 0.0001	Step 0: matched cached baseline	All Steps 0–20; Step20 peak	Starting global checkpoint only; no adapted query states	80.758% @ step 20; final 80.758% @ step 20
Head-only TTT from original reference — SGD 0.000001	Step 0: matched cached baseline	All Steps 0–20; Step17 peak; Step20 final	Starting global checkpoint only; no adapted query states	80.590% @ step 17; final 80.590% @ step 20
Head-only TTT from CE20 — Adam 0.01	Step 0: matched cached baseline	All Steps 0–50; Step0 peak; Step50 final	Starting global checkpoint only; no adapted query states	82.084% before → 81.098% final (step 50); no adapted step wins
Shared row identities and labels	Applies to validation-logit rows; embeddings also carry split-specific IDs/labels	50,000 sample IDs; 50,000 targets; Format	—	Every validation logit array is 50,000 × 1,000 FP32, not probabilities; embeddings are separately indexed
Model, loss and hyperparameter details	Logits archive	Complete methods	Checkpoints archive; Same methods	Architecture, fitted head, parent/teacher identities, optimizer, masking, preprocessing and TTT
Full results, figures and explanations	Published fixed-endpoint and paired-logit analyses	Read the full report; Offline HTML	Evidence ZIP	Includes correction/harm, calibration, saliency and logit-overlap evidence
Machine-readable indexes, coverage and provenance	Coverage / missing files	Training CSV; TTT CSV; Logit provenance	Checkpoint provenance; Retention contract	Hashes stay in provenance; old artifact paths remain valid
Engineering smoke/canary checkpoints — not scientific results	Not extra experimental arms	No production-logit comparison	Engineering-only index	Excluded from best-result rankings
Other later TTT output studies — not published here	Normalized CE20 / CE+SD head TTT and raw-feature paper600	No download in this release; Explicit exclusions	No implied publication	Do not mistake the older SGD50 / Adam60 rows for these later runs

Embedding coverage — final versus selected checkpoints

Encoder checkpoint	Support pool	Validation	Development
Original pretrained/reference backbone	1,231,154 × 768	50,000 × 768	Not exported
CE-only — final epoch 20	1,231,154 × 768	50,000 × 768	5,000 × 768
CE+SD original teacher — selected epoch 10	1,231,154 × 768	50,000 × 768	5,000 × 768
CE+SD original teacher — final epoch 20	1,231,154 × 768	50,000 × 768	Not exported
CE10 → continued CE, repeated schedule — final epoch 20	1,231,154 × 768	50,000 × 768	Not exported
CE10 → CE+SD, teacher CE10 — final epoch 20	1,231,154 × 768	50,000 × 768	Not exported
CE+SD10 → CE+SD, teacher SD10 — final epoch 20	1,231,154 × 768	50,000 × 768	Not exported

16 published embedding arrays across seven named encoder states. All five main epoch-20 endpoints now have training-support and validation embeddings. All IDs, labels and explanations.

Embedding guide: Checkpoint-by-split downloads and an executable embeddings + logits example. October TTT: All eight raw-head SGD/Muon results and exact recipes. Both now use readable paths; old content-addressed stores remain unchanged.

What the arm names mean

An arm is one experimental configuration, not an individual image. CE is ordinary supervised cross-entropy; CE+SD adds learning from a frozen teacher, with exactly 0.5 × clean CE + 0.5 × masked forward KL. An original teacher is the shared fitted epoch-0 model. A refreshed teacher is the named CE10 or CE+SD10 model, frozen for that branch.

Fast5 → floor means cosine decay from LR 0.005 to that floor over five epochs, then hold. Cosine10 instead decays for ten epochs. Constant LR never decays. The repeated Fast5 branches restart the schedule phase after epoch 10 while retaining the parent's optimizer state. Their “before” links therefore point to CE10 or CE+SD10, not epoch 0.

TTT adapts only a fresh linear-head copy for each query, using labelled training neighbours; the backbone and master head stay frozen. Its numbers are update steps, not global training epochs. Use the row's cached step-0 baseline: original-reference TTT starts at 80.588%, and CE20 TTT at 82.084%. These are not bit-identical to direct-image epoch 0 (80.586%) or CE20 (82.086%).

A harm is different from an arm: the baseline was correct and the candidate becomes wrong. A correction is baseline-wrong → candidate-correct. Neither count alone establishes a net improvement; the baseline must be named.

Reading the table correctly

(best) among global-training checkpoints: the highest observed validation Top-1 is 82.634%, original-teacher CE+SD / Fast5 → 1e-4 / epoch 10. It is not the last checkpoint.
(best) among CE controls: 82.170%, CE / Fast5 → 1e-4 / epoch 14. The original fixed epoch-20 CE result remains 82.086%; Adam TTT still starts from CE20, not CE14.
(best) in the October raw-head TTT fixed-final grid: CE+SD10 + Muon LR 0.005 reaches 83.234% at step 20. Its NLL improves but ECE worsens; this is already-open-validation, single-seed evidence, not fresh confirmation.
The original-reference SGD TTT best label is within its own original-reference SGD grid only, not an overall win. Adam60 is a negative result.
“Best” is descriptive, validation-informed, single-seed selection, not untouched-test evidence or a claim of superiority on every metric. Peak and fixed/final endpoints are displayed separately.
All original CE/CE+SD rows share one fitted epoch-0 reference: pretrained backbone plus already-fitted ImageNet head. The legacy /ce/ reference folder also applies to CE+SD. Its checkpoint's head-fit epoch 10 is global fine-tuning epoch 0.
Original-teacher CE+SD at 1e-4 has every checkpoint + optimizer from E1–20, but raw logits only at E1–10 and E15–20. E11–14 scalar results do not substitute for missing arrays.
Click the arm name or All link for a plain-language description and one row per epoch/step. Every global-training arm page includes its correct before row; every TTT page keeps its global parent separate from unretained adapted state.

Access and file semantics

Cards are public; downloads require separate manual approval on each repository. Existing permissions, licensing, payload bytes and old links are unchanged. Embedding files instead contain raw pooled 768-dimensional FP16 features and split-specific IDs/labels; their headers, not two incorrect legacy dtype descriptions, determine precision. Each logits file contains all validation images and all classes; pair it with the shared IDs/targets from the table. The model file bundles retained global model/optimizer state; a TTT parent link is not a per-query adapted model archive.

Hashes are provenance, not the reader-facing experiment names. The new navigation follows main; actual array/checkpoint links pin the existing immutable payload revisions. Avoid materializing both legacy and readable layouts when downloading the archive.

Downloads last month: 98

Total file size:

155 GB