DINOv3 experiments — results, checkpoints and optimizers
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 the overall (best) result and 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.
| 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) |
Original-teacher CE+SD — main runFast5 → 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) — overall; final 82.624% @ epoch 20 |
CE-only — main supervised controlFast5 → 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 floorFast5 → 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 floorFast5 → 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 floorFast5 → 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 floorFast5 → 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 scheduleRepeat 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 CE10Repeat 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+SD10Repeat 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 every published array | 50,000 sample IDs; 50,000 targets; Format | — | Every array is 50,000 × 1,000 FP32 raw logits, not probabilities |
| 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 |
| Later TTT 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 |
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) overall: 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. - The 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-4has 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. 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.