EP heads for 36 frozen encoders (12 best-epoch, 24 final-epoch with dual metadata)

README.md

@@ -0,0 +1,91 @@
+---
+license: apache-2.0
+tags:
+- image-classification
+- probing
+- frozen-features
+- imagenet-1k
+library_name: pytorch
+---
+
+# Efficient Probing heads
+
+Trained EP (efficient probing) heads for 36 frozen encoders, from the standing
+ImageNet-1k benchmark at [billpsomas/efficient-probing](https://github.com/billpsomas/efficient-probing)
+(ICLR 2026). Each head is a multi-query cross-attention pooling + BatchNorm + linear
+classifier, trained for 90 epochs with LARS on frozen features. **No backbone weights
+are included** -- each `config.json` records the exact `--model` / loader flags to
+rebuild the encoder from its original source.
+
+Every head ships with its accuracy **at the saved epoch** in `meta`/`config.json`:
+12 heads are the run's peak (they reproduce the leaderboard number exactly); the rest
+are final-epoch heads, with both figures recorded so nothing is ambiguous. DiT-XL/2
+is absent (no surviving checkpoint; see the manifest).
+
+## Loading
+
+```python
+from huggingface_hub import hf_hub_download
+import torch
+
+path = hf_hub_download("billpsomas/efficient-probing-heads",
+                       "dinov3_vit7b/ep_head.pth")
+ck = torch.load(path, map_location="cpu", weights_only=False)
+head_state, meta = ck["state_dict"], ck["meta"]
+# with the benchmark repo on PYTHONPATH:
+#   model = backbones.build_backbone(args, device)   # args from meta
+#   probe_heads.build_probe_head(model, args)
+#   model.head.load_state_dict(head_state, strict=True)
+```
+
+Or evaluate directly with the benchmark's tool:
+
+```
+python tools/eval_reimagenet.py predict <backbone flags from config.json> \
+    --head_ckpt ep_head.pth --pred_out preds.json
+```
+
+## Heads
+
+| encoder | EP variant | top-1 @ saved epoch | epoch | checkpoint is |
+|---|---|---:|---:|---|
+| DINOv3 ViT-7B/16 | ep_all | 88.36 | 6 | peak |
+| MetaCLIP2 ViT-bigG/14-378 | ep | 88.12 | 6 | peak |
+| EVA02-CLIP E-14-plus | ep | 87.98 | 6 | peak |
+| EVA02-CLIP E-14 | ep | 87.70 | 6 | peak |
+| SigLIP2 SO400M/14 | ep | 85.64 | 29 | final epoch |
+| PE-Core L-14/336 | ep | 87.25 | 12 | peak |
+| MetaCLIP2 ViT-bigG/14 | ep | 87.11 | 6 | peak |
+| SigLIP2 ViT-L/16 | ep | 87.06 | 6 | peak |
+| DINOv3 ViT-L/16 | ep_all | 86.73 | 19 | final epoch |
+| AIMv2 ViT-L/14 | ep | 85.62 | 19 | final epoch |
+| SigLIP ViT-L/16 | ep | 85.93 | 6 | peak |
+| DINOv2 ViT-L/14 | ep_all | 85.56 | 15 | peak |
+| Franca ViT-L/14 | ep_all | 84.28 | 14 | peak |
+| DINOv3 ViT-B/16 | ep_all | 83.77 | 20 | final epoch |
+| DINOv2 ViT-B/14 | ep | 83.61 | 25 | final epoch |
+| RADIO ViT-L/16 | ep | 83.40 | 89 | final epoch |
+| EVA02 ViT-L/14 | ep | 83.22 | 89 | final epoch |
+| CLIP ViT-L/14 | ep | 83.22 | 11 | peak |
+| CAPI ViT-L/14 | ep | 82.43 | 89 | final epoch |
+| BEiTv2 ViT-B/16 | ep | 81.32 | 89 | final epoch |
+| RADIO ViT-B/16 | ep | 80.26 | 89 | final epoch |
+| iBOT ViT-L/16 | ep_all | 79.43 | 89 | final epoch |
+| Hiera ViT-H/16 | ep | 79.82 | 89 | final epoch |
+| MAE ViT-L/16 | ep | 79.43 | 89 | final epoch |
+| I-JEPA ViT-H/14 | ep | 78.80 | 89 | final epoch |
+| iBOT ViT-B/16 | ep_all | 78.62 | 89 | final epoch |
+| Hiera ViT-L/16 | ep | 78.51 | 83 | final epoch |
+| CLIP ViT-B/16 | ep_all | 77.85 | 11 | peak |
+| DINO ViT-B/16 | ep_all | 77.08 | 89 | final epoch |
+| MoCov3 ViT-B/16 | ep_all | 76.21 | 89 | final epoch |
+| Hiera ViT-B/16 | ep | 75.63 | 88 | final epoch |
+| MAE ViT-B/16 | ep | 75.35 | 86 | final epoch |
+| MaskFeat ViT-B/16 | ep | 71.68 | 89 | final epoch |
+| MaskFeat ViT-L/16 | ep | 69.56 | 89 | final epoch |
+| SimMIM ViT-B/16 | ep | 64.81 | 89 | final epoch |
+| MAE ViT-S/16 | ep | 64.56 | 89 | final epoch |
+
+Full provenance (training logs, exact commands, the leaderboard itself) lives in the
+[GitHub repo](https://github.com/billpsomas/efficient-probing). Heads were trained on
+ImageNet-1k; use accordingly.

aimv2_vitl/config.json

@@ -0,0 +1,19 @@
+{
+ "method": "AIMv2",
+ "arch": "ViT-L/14",
+ "pretrain": "DFN-2B + COYO + HQITP",
+ "image_size": 224,
+ "model": "aimv2-large-patch14-224",
+ "loader": "aimv2",
+ "pretrain_tag": "",
+ "finetune": "",
+ "extra_flags": "",
+ "cls_features": "ep",
+ "ep_queries": 32,
+ "d_out": 1,
+ "head_epoch": 19,
+ "val_acc1_at_head_epoch": 85.62,
+ "table_best_acc1": 85.94,
+ "note": "final-epoch head, not best-epoch: checkpoints overwrite per epoch. table_best_acc1 is the peak of the same run; the difference is the post-peak decline of a frozen-feature probe, largest on VLM encoders.",
+ "training_log": "logs/aimv2_vitl/ep.txt"
+}

aimv2_vitl/ep_head.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f7565e9ccc7791a3cdcb2ae6cffd2fa2650f2d994807a6b5415de875e4d9bc41
+size 8437201

beitv2_vitb/config.json

@@ -0,0 +1,19 @@
+{
+ "method": "BEiTv2",
+ "arch": "ViT-B/16",
+ "pretrain": "IN-1K",
+ "image_size": 224,
+ "model": "vit_base_patch16",
+ "loader": "simmim",
+ "pretrain_tag": "",
+ "finetune": "vit_base_patch16_224.beitv2/checkpoint-799.pth",
+ "extra_flags": "",
+ "cls_features": "ep",
+ "ep_queries": 32,
+ "d_out": 1,
+ "head_epoch": 89,
+ "val_acc1_at_head_epoch": 81.32,
+ "table_best_acc1": 81.39,
+ "note": "final-epoch head, not best-epoch: checkpoints overwrite per epoch. table_best_acc1 is the peak of the same run; the difference is the post-peak decline of a frozen-feature probe, largest on VLM encoders.",
+ "training_log": "logs/beitv2_vitb/ep.txt"
+}

beitv2_vitb/ep_head.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:98ff601343eb6f603dac0da5171a19871890274c1d3e9b76b28fae7951445250
+size 5543441

capi_vitl/config.json

@@ -0,0 +1,19 @@
+{
+ "method": "CAPI",
+ "arch": "ViT-L/14",
+ "pretrain": "IN-1K",
+ "image_size": 224,
+ "model": "capi_vitl14_in1k",
+ "loader": "capi",
+ "pretrain_tag": "",
+ "finetune": "",
+ "extra_flags": "",
+ "cls_features": "ep",
+ "ep_queries": 32,
+ "d_out": 1,
+ "head_epoch": 89,
+ "val_acc1_at_head_epoch": 82.43,
+ "table_best_acc1": 82.98,
+ "note": "final-epoch head, not best-epoch: checkpoints overwrite per epoch. table_best_acc1 is the peak of the same run; the difference is the post-peak decline of a frozen-feature probe, largest on VLM encoders.",
+ "training_log": "logs/capi_vitl/ep.txt"
+}

capi_vitl/ep_head.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ce826a1e518624a416e0a8887c4532170a271d069bf1176287d9bca633bc58f6
+size 8437201

clip_vitb16/config.json

@@ -0,0 +1,19 @@
+{
+ "method": "CLIP",
+ "arch": "ViT-B/16",
+ "pretrain": "WIT-400M",
+ "image_size": 224,
+ "model": "ViT-B-16",
+ "loader": "openclip",
+ "pretrain_tag": "openai",
+ "finetune": "",
+ "extra_flags": "",
+ "cls_features": "ep_all",
+ "ep_queries": 32,
+ "d_out": 1,
+ "head_epoch": 11,
+ "val_acc1_at_head_epoch": 77.85,
+ "table_best_acc1": 77.85,
+ "note": "best-epoch head: saved at the run's peak, the number the table reports.",
+ "training_log": "logs/clip_vitb16/ep.txt"
+}

clip_vitb16/ep_head.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0ccc806165e35b6b2bf24e5d969168700f1445a8fd9b4a75b54ecf106bc87310
+size 3173713

clip_vitl/config.json

@@ -0,0 +1,19 @@
+{
+ "method": "CLIP",
+ "arch": "ViT-L/14",
+ "pretrain": "WIT-400M",
+ "image_size": 224,
+ "model": "ViT-L-14",
+ "loader": "openclip",
+ "pretrain_tag": "openai",
+ "finetune": "",
+ "extra_flags": "",
+ "cls_features": "ep",
+ "ep_queries": 32,
+ "d_out": 1,
+ "head_epoch": 11,
+ "val_acc1_at_head_epoch": 83.22,
+ "table_best_acc1": 83.22,
+ "note": "best-epoch head: saved at the run's peak, the number the table reports.",
+ "training_log": "logs/clip_vitl/ep.txt"
+}

clip_vitl/ep_head.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1bd6460e5b20f6b7355a8522e7034d195a6000127dd6b9af8591cde2ef0b7515
+size 8437073

dino_vitb/config.json

@@ -0,0 +1,19 @@
+{
+ "method": "DINO",
+ "arch": "ViT-B/16",
+ "pretrain": "IN-1K",
+ "image_size": 224,
+ "model": "vit_base_patch16",
+ "loader": "",
+ "pretrain_tag": "",
+ "finetune": "vit_base_patch16_224.dino",
+ "extra_flags": "",
+ "cls_features": "ep_all",
+ "ep_queries": 32,
+ "d_out": 1,
+ "head_epoch": 89,
+ "val_acc1_at_head_epoch": 77.08,
+ "table_best_acc1": 77.39,
+ "note": "final-epoch head, not best-epoch: checkpoints overwrite per epoch. table_best_acc1 is the peak of the same run; the difference is the post-peak decline of a frozen-feature probe, largest on VLM encoders.",
+ "training_log": "logs/dino_vitb/ep.txt"
+}

dino_vitb/ep_head.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:76941d4fb977f8b2c21b998c805ef37935886c778b0ad12b6a7c1edd904cce10
+size 5543377

dinov2_vitb/config.json

@@ -0,0 +1,19 @@
+{
+ "method": "DINOv2",
+ "arch": "ViT-B/14",
+ "pretrain": "LVD-142M",
+ "image_size": 224,
+ "model": "dinov2_vitb14",
+ "loader": "hub",
+ "pretrain_tag": "",
+ "finetune": "",
+ "extra_flags": "",
+ "cls_features": "ep",
+ "ep_queries": 32,
+ "d_out": 1,
+ "head_epoch": 25,
+ "val_acc1_at_head_epoch": 83.61,
+ "table_best_acc1": 84.01,
+ "note": "final-epoch head, not best-epoch: checkpoints overwrite per epoch. table_best_acc1 is the peak of the same run; the difference is the post-peak decline of a frozen-feature probe, largest on VLM encoders.",
+ "training_log": "logs/dinov2_vitb/ep.txt"
+}

dinov2_vitb/ep_head.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:13bcf74040d8d47cfca8f7eea7ef80b8391ab14bbd0fe1a1f94b4ad7b1687668
+size 5543377

dinov2_vitl/config.json

@@ -0,0 +1,19 @@
+{
+ "method": "DINOv2",
+ "arch": "ViT-L/14",
+ "pretrain": "LVD-142M",
+ "image_size": 224,
+ "model": "dinov2_vitl14",
+ "loader": "hub",
+ "pretrain_tag": "",
+ "finetune": "",
+ "extra_flags": "",
+ "cls_features": "ep_all",
+ "ep_queries": 32,
+ "d_out": 1,
+ "head_epoch": 15,
+ "val_acc1_at_head_epoch": 85.56,
+ "table_best_acc1": 85.56,
+ "note": "best-epoch head: saved at the run's peak, the number the table reports.",
+ "training_log": "logs/dinov2_vitl/ep.txt"
+}

dinov2_vitl/ep_head.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6d72ad24de49a8af0eddfc062203a32f4536a5a2871ccf9fad5c6f0c610e868e
+size 8437073

dinov3_vit7b/config.json

@@ -0,0 +1,19 @@
+{
+ "method": "DINOv3",
+ "arch": "ViT-7B/16",
+ "pretrain": "LVD-1689M",
+ "image_size": 224,
+ "model": "dinov3_vit7b16",
+ "loader": "dinov3",
+ "pretrain_tag": "",
+ "finetune": "",
+ "extra_flags": "",
+ "cls_features": "ep_all",
+ "ep_queries": 32,
+ "d_out": 1,
+ "head_epoch": 6,
+ "val_acc1_at_head_epoch": 88.36,
+ "table_best_acc1": 88.36,
+ "note": "best-epoch head: saved at the run's peak, the number the table reports.",
+ "training_log": "logs/dinov3_vit7b/ep.txt"
+}

dinov3_vit7b/ep_head.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7687980bc0cba8f25e9157bfb8279efc4f67487ef787c3a36c5995a133add666
+size 84057425

dinov3_vitb/config.json

@@ -0,0 +1,19 @@
+{
+ "method": "DINOv3",
+ "arch": "ViT-B/16",
+ "pretrain": "LVD-1689M",
+ "image_size": 224,
+ "model": "dinov3_vitb16",
+ "loader": "dinov3",
+ "pretrain_tag": "",
+ "finetune": "",
+ "extra_flags": "",
+ "cls_features": "ep_all",
+ "ep_queries": 32,
+ "d_out": 1,
+ "head_epoch": 20,
+ "val_acc1_at_head_epoch": 83.77,
+ "table_best_acc1": 84.11,
+ "note": "final-epoch head, not best-epoch: checkpoints overwrite per epoch. table_best_acc1 is the peak of the same run; the difference is the post-peak decline of a frozen-feature probe, largest on VLM encoders.",
+ "training_log": "logs/dinov3_vitb/ep.txt"
+}

dinov3_vitb/ep_head.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a55d603fefc8eec3bc1fb6ecca75126ee6593303b5d382e2dd72e8a6fc57245a
+size 5543377

dinov3_vitl/config.json

@@ -0,0 +1,19 @@
+{
+ "method": "DINOv3",
+ "arch": "ViT-L/16",
+ "pretrain": "LVD-1689M",
+ "image_size": 224,
+ "model": "dinov3_vitl16",
+ "loader": "dinov3",
+ "pretrain_tag": "",
+ "finetune": "",
+ "extra_flags": "",
+ "cls_features": "ep_all",
+ "ep_queries": 32,
+ "d_out": 1,
+ "head_epoch": 19,
+ "val_acc1_at_head_epoch": 86.73,
+ "table_best_acc1": 87.05,
+ "note": "final-epoch head, not best-epoch: checkpoints overwrite per epoch. table_best_acc1 is the peak of the same run; the difference is the post-peak decline of a frozen-feature probe, largest on VLM encoders.",
+ "training_log": "logs/dinov3_vitl/ep.txt"
+}

dinov3_vitl/ep_head.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:718a1ac2802741c31199265fa8e501a40cf91eaf156e269837d8a0ba5ffc6040
+size 8437201

eva02_e14/config.json

@@ -0,0 +1,19 @@
+{
+ "method": "EVA02-CLIP",
+ "arch": "E-14",
+ "pretrain": "LAION-2B (4B seen)",
+ "image_size": 224,
+ "model": "EVA02-E-14",
+ "loader": "openclip",
+ "pretrain_tag": "laion2b_s4b_b115k",
+ "finetune": "",
+ "extra_flags": "",
+ "cls_features": "ep",
+ "ep_queries": 32,
+ "d_out": 1,
+ "head_epoch": 6,
+ "val_acc1_at_head_epoch": 87.7,
+ "table_best_acc1": 87.7,
+ "note": "best-epoch head: saved at the run's peak, the number the table reports.",
+ "training_log": "logs/eva02_e14/ep.txt"
+}

eva02_e14/ep_head.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5bc9dd3fa66bc9c7a5a38287d02adf004765544aa935dc3b16aeab1da1456b68
+size 20264273

eva02_e14_plus/config.json

@@ -0,0 +1,19 @@
+{
+ "method": "EVA02-CLIP",
+ "arch": "E-14-plus",
+ "pretrain": "LAION-2B (9B seen)",
+ "image_size": 224,
+ "model": "EVA02-E-14-plus",
+ "loader": "openclip",
+ "pretrain_tag": "laion2b_s9b_b144k",
+ "finetune": "",
+ "extra_flags": "",
+ "cls_features": "ep",
+ "ep_queries": 32,
+ "d_out": 1,
+ "head_epoch": 6,
+ "val_acc1_at_head_epoch": 87.98,
+ "table_best_acc1": 87.98,
+ "note": "best-epoch head: saved at the run's peak, the number the table reports.",
+ "training_log": "logs/eva02_e14_plus/ep.txt"
+}

eva02_e14_plus/ep_head.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:868d4b18ed28acfab9871585d9eeb7ce61d56b60589e378f9dddc0b018127062
+size 20264337

eva02_vitl_mim/config.json

@@ -0,0 +1,19 @@
+{
+ "method": "EVA02",
+ "arch": "ViT-L/14",
+ "pretrain": "IN-22K",
+ "image_size": 224,
+ "model": "eva02_large_patch14_224.mim_in22k",
+ "loader": "timm",
+ "pretrain_tag": "",
+ "finetune": "",
+ "extra_flags": "",
+ "cls_features": "ep",
+ "ep_queries": 32,
+ "d_out": 1,
+ "head_epoch": 89,
+ "val_acc1_at_head_epoch": 83.22,
+ "table_best_acc1": 83.58,
+ "note": "final-epoch head, not best-epoch: checkpoints overwrite per epoch. table_best_acc1 is the peak of the same run; the difference is the post-peak decline of a frozen-feature probe, largest on VLM encoders.",
+ "training_log": "logs/eva02_vitl_mim/ep.txt"
+}

eva02_vitl_mim/ep_head.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:94b945f2cf9c040e65cac57bffe8a79b211e6dec6b889f507423b591897cd09d
+size 8437201

franca_laion/config.json

@@ -0,0 +1,19 @@
+{
+ "method": "Franca",
+ "arch": "ViT-L/14",
+ "pretrain": "LAION-600M",
+ "image_size": 224,
+ "model": "franca_vitl14",
+ "loader": "franca",
+ "pretrain_tag": "",
+ "finetune": "",
+ "extra_flags": "--franca_weights LAION",
+ "cls_features": "ep_all",
+ "ep_queries": 32,
+ "d_out": 1,
+ "head_epoch": 14,
+ "val_acc1_at_head_epoch": 84.28,
+ "table_best_acc1": 84.28,
+ "note": "best-epoch head: saved at the run's peak, the number the table reports.",
+ "training_log": "logs/franca_laion/ep_all.txt"
+}

franca_laion/ep_head.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1f930b15a1f4a11f723c943b03e2bc83283730e4619c792fcec6e3c07af6eea6
+size 8437073

hiera_base/config.json

@@ -0,0 +1,19 @@
+{
+ "method": "Hiera",
+ "arch": "ViT-B/16",
+ "pretrain": "IN-1K",
+ "image_size": 224,
+ "model": "hiera_base_224.mae",
+ "loader": "timm",
+ "pretrain_tag": "",
+ "finetune": "",
+ "extra_flags": "",
+ "cls_features": "ep",
+ "ep_queries": 32,
+ "d_out": 1,
+ "head_epoch": 88,
+ "val_acc1_at_head_epoch": 75.63,
+ "table_best_acc1": 75.68,
+ "note": "final-epoch head, not best-epoch: checkpoints overwrite per epoch. table_best_acc1 is the peak of the same run; the difference is the post-peak decline of a frozen-feature probe, largest on VLM encoders.",
+ "training_log": "logs/hiera_base/ep.txt"
+}

hiera_base/ep_head.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c023a84619b4bd5b22f043e9ca4d14afbcb608975f23eb7d9c6a676317304d2d
+size 5543377

hiera_huge/config.json

@@ -0,0 +1,19 @@
+{
+ "method": "Hiera",
+ "arch": "ViT-H/16",
+ "pretrain": "IN-1K",
+ "image_size": 224,
+ "model": "hiera_huge_224.mae",
+ "loader": "timm",
+ "pretrain_tag": "",
+ "finetune": "",
+ "extra_flags": "",
+ "cls_features": "ep",
+ "ep_queries": 32,
+ "d_out": 1,
+ "head_epoch": 89,
+ "val_acc1_at_head_epoch": 79.82,
+ "table_best_acc1": 79.89,
+ "note": "final-epoch head, not best-epoch: checkpoints overwrite per epoch. table_best_acc1 is the peak of the same run; the difference is the post-peak decline of a frozen-feature probe, largest on VLM encoders.",
+ "training_log": "logs/hiera_huge/ep.txt"
+}

hiera_huge/ep_head.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:183f0f8dcb99bc13576410a06968c6a9040a1ce2ce2c6b931141bcd01dbc9836
+size 25255377

hiera_large/config.json

@@ -0,0 +1,19 @@
+{
+ "method": "Hiera",
+ "arch": "ViT-L/16",
+ "pretrain": "IN-1K",
+ "image_size": 224,
+ "model": "hiera_large_224.mae",
+ "loader": "timm",
+ "pretrain_tag": "",
+ "finetune": "",
+ "extra_flags": "",
+ "cls_features": "ep",
+ "ep_queries": 32,
+ "d_out": 1,
+ "head_epoch": 83,
+ "val_acc1_at_head_epoch": 78.51,
+ "table_best_acc1": 78.55,
+ "note": "final-epoch head, not best-epoch: checkpoints overwrite per epoch. table_best_acc1 is the peak of the same run; the difference is the post-peak decline of a frozen-feature probe, largest on VLM encoders.",
+ "training_log": "logs/hiera_large/ep.txt"
+}

hiera_large/ep_head.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ffd14707b8652a2725a6788cb57b8b2c3ef04d7c29f501a49048ae13eb0e8f03
+size 10080721

ibot_vitb/config.json

@@ -0,0 +1,19 @@
+{
+ "method": "iBOT",
+ "arch": "ViT-B/16",
+ "pretrain": "IN-1K",
+ "image_size": 224,
+ "model": "vit_base_patch16",
+ "loader": "",
+ "pretrain_tag": "",
+ "finetune": "vit_base_patch16_224.ibot/checkpoint-799.pth",
+ "extra_flags": "",
+ "cls_features": "ep_all",
+ "ep_queries": 32,
+ "d_out": 1,
+ "head_epoch": 89,
+ "val_acc1_at_head_epoch": 78.62,
+ "table_best_acc1": 78.98,
+ "note": "final-epoch head, not best-epoch: checkpoints overwrite per epoch. table_best_acc1 is the peak of the same run; the difference is the post-peak decline of a frozen-feature probe, largest on VLM encoders.",
+ "training_log": "logs/ibot_vitb/ep.txt"
+}

ibot_vitb/ep_head.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b21c07f716b4aa0c0274b64478e2e98ebb412597206627321cca81ae685c363d
+size 5543441

ibot_vitl/config.json

@@ -0,0 +1,19 @@
+{
+ "method": "iBOT",
+ "arch": "ViT-L/16",
+ "pretrain": "IN-1K",
+ "image_size": 224,
+ "model": "vit_large_patch16",
+ "loader": "",
+ "pretrain_tag": "",
+ "finetune": "vit_large_patch16_224.ibot/checkpoint-799.pth",
+ "extra_flags": "",
+ "cls_features": "ep_all",
+ "ep_queries": 32,
+ "d_out": 1,
+ "head_epoch": 89,
+ "val_acc1_at_head_epoch": 79.43,
+ "table_best_acc1": 79.97,
+ "note": "final-epoch head, not best-epoch: checkpoints overwrite per epoch. table_best_acc1 is the peak of the same run; the difference is the post-peak decline of a frozen-feature probe, largest on VLM encoders.",
+ "training_log": "logs/ibot_vitl/ep.txt"
+}

ibot_vitl/ep_head.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:839dcfeabe8eff1868d73e6d4111d6ba6c6496f19e66cc3c8bc2949110c2206f
+size 8437265

ijepa_vith/config.json

@@ -0,0 +1,19 @@
+{
+ "method": "I-JEPA",
+ "arch": "ViT-H/14",
+ "pretrain": "IN-1K",
+ "image_size": 224,
+ "model": "vit_huge_patch14",
+ "loader": "",
+ "pretrain_tag": "",
+ "finetune": "vit_huge_patch14_224.ijepa/checkpoint-799.pth",
+ "extra_flags": "--no_cls_token",
+ "cls_features": "ep",
+ "ep_queries": 32,
+ "d_out": 1,
+ "head_epoch": 89,
+ "val_acc1_at_head_epoch": 78.8,
+ "table_best_acc1": 79.0,
+ "note": "final-epoch head, not best-epoch: checkpoints overwrite per epoch. table_best_acc1 is the peak of the same run; the difference is the post-peak decline of a frozen-feature probe, largest on VLM encoders.",
+ "training_log": "logs/ijepa_vith/ep.txt"
+}

ijepa_vith/ep_head.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d1d74dcc70511357a6911785ddb7983a3f19368a1a497b62c618b80260cad966
+size 11855377

mae_base/config.json

@@ -0,0 +1,19 @@
+{
+ "method": "MAE",
+ "arch": "ViT-B/16",
+ "pretrain": "IN-1K",
+ "image_size": 224,
+ "model": "vit_base_patch16",
+ "loader": "",
+ "pretrain_tag": "",
+ "finetune": "vit_base_patch16_224.mae",
+ "extra_flags": "",
+ "cls_features": "ep",
+ "ep_queries": 32,
+ "d_out": 1,
+ "head_epoch": 86,
+ "val_acc1_at_head_epoch": 75.35,
+ "table_best_acc1": 75.46,
+ "note": "final-epoch head, not best-epoch: checkpoints overwrite per epoch. table_best_acc1 is the peak of the same run; the difference is the post-peak decline of a frozen-feature probe, largest on VLM encoders.",
+ "training_log": "logs/mae_base/ep.txt"
+}

mae_base/ep_head.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9425963f1a1575305694ca1875bed9f80140e53cd263548cd81c6d68fbacb1c7
+size 5543377

mae_vitl/config.json

@@ -0,0 +1,19 @@
+{
+ "method": "MAE",
+ "arch": "ViT-L/16",
+ "pretrain": "IN-1K",
+ "image_size": 224,
+ "model": "vit_large_patch16",
+ "loader": "",
+ "pretrain_tag": "",
+ "finetune": "vit_large_patch16_224.mae",
+ "extra_flags": "",
+ "cls_features": "ep",
+ "ep_queries": 32,
+ "d_out": 1,
+ "head_epoch": 89,
+ "val_acc1_at_head_epoch": 79.43,
+ "table_best_acc1": 79.52,
+ "note": "final-epoch head, not best-epoch: checkpoints overwrite per epoch. table_best_acc1 is the peak of the same run; the difference is the post-peak decline of a frozen-feature probe, largest on VLM encoders.",
+ "training_log": "logs/mae_vitl/ep.txt"
+}

mae_vitl/ep_head.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c3935dbcf3cf125e6d2c4f6de070bac04c250c0fa15bb493c6d12d1e3ea6e151
+size 8437201

mae_vits/config.json

@@ -0,0 +1,19 @@
+{
+ "method": "MAE",
+ "arch": "ViT-S/16",
+ "pretrain": "IN-1K",
+ "image_size": 224,
+ "model": "vit_small_patch16",
+ "loader": "",
+ "pretrain_tag": "",
+ "finetune": "vit_small_patch16_224.mae/checkpoint-799.pth",
+ "extra_flags": "",
+ "cls_features": "ep",
+ "ep_queries": 32,
+ "d_out": 1,
+ "head_epoch": 89,
+ "val_acc1_at_head_epoch": 64.56,
+ "table_best_acc1": 64.58,
+ "note": "final-epoch head, not best-epoch: checkpoints overwrite per epoch. table_best_acc1 is the peak of the same run; the difference is the post-peak decline of a frozen-feature probe, largest on VLM encoders.",
+ "training_log": "logs/mae_vits/ep.txt"
+}

mae_vits/ep_head.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8ba33f00307a72367c0a560dc36b364a24f0a15f2f55b8c1a7120c1d0dcadf1b
+size 2185681

manifest.json

@@ -0,0 +1,803 @@
+{
+ "heads": [
+  {
+   "method": "MAE",
+   "arch": "ViT-S/16",
+   "pretrain": "IN-1K",
+   "image_size": 224,
+   "model": "vit_small_patch16",
+   "loader": "",
+   "pretrain_tag": "",
+   "finetune": "vit_small_patch16_224.mae/checkpoint-799.pth",
+   "extra_flags": "",
+   "cls_features": "ep",
+   "ep_queries": 32,
+   "d_out": 1,
+   "head_epoch": 89,
+   "val_acc1_at_head_epoch": 64.56,
+   "table_best_acc1": 64.58,
+   "note": "final-epoch head, not best-epoch: checkpoints overwrite per epoch. table_best_acc1 is the peak of the same run; the difference is the post-peak decline of a frozen-feature probe, largest on VLM encoders.",
+   "training_log": "logs/mae_vits/ep.txt",
+   "params_incl_bn_stats": 545513,
+   "file": "mae_vits/ep_head.pth",
+   "size_mb": 2.2
+  },
+  {
+   "method": "MAE",
+   "arch": "ViT-B/16",
+   "pretrain": "IN-1K",
+   "image_size": 224,
+   "model": "vit_base_patch16",
+   "loader": "",
+   "pretrain_tag": "",
+   "finetune": "vit_base_patch16_224.mae",
+   "extra_flags": "",
+   "cls_features": "ep",
+   "ep_queries": 32,
+   "d_out": 1,
+   "head_epoch": 86,
+   "val_acc1_at_head_epoch": 75.35,
+   "table_best_acc1": 75.46,
+   "note": "final-epoch head, not best-epoch: checkpoints overwrite per epoch. table_best_acc1 is the peak of the same run; the difference is the post-peak decline of a frozen-feature probe, largest on VLM encoders.",
+   "training_log": "logs/mae_base/ep.txt",
+   "params_incl_bn_stats": 1384937,
+   "file": "mae_base/ep_head.pth",
+   "size_mb": 5.5
+  },
+  {
+   "method": "MAE",
+   "arch": "ViT-L/16",
+   "pretrain": "IN-1K",
+   "image_size": 224,
+   "model": "vit_large_patch16",
+   "loader": "",
+   "pretrain_tag": "",
+   "finetune": "vit_large_patch16_224.mae",
+   "extra_flags": "",
+   "cls_features": "ep",
+   "ep_queries": 32,
+   "d_out": 1,
+   "head_epoch": 89,
+   "val_acc1_at_head_epoch": 79.43,
+   "table_best_acc1": 79.52,
+   "note": "final-epoch head, not best-epoch: checkpoints overwrite per epoch. table_best_acc1 is the peak of the same run; the difference is the post-peak decline of a frozen-feature probe, largest on VLM encoders.",
+   "training_log": "logs/mae_vitl/ep.txt",
+   "params_incl_bn_stats": 2108393,
+   "file": "mae_vitl/ep_head.pth",
+   "size_mb": 8.4
+  },
+  {
+   "method": "BEiTv2",
+   "arch": "ViT-B/16",
+   "pretrain": "IN-1K",
+   "image_size": 224,
+   "model": "vit_base_patch16",
+   "loader": "simmim",
+   "pretrain_tag": "",
+   "finetune": "vit_base_patch16_224.beitv2/checkpoint-799.pth",
+   "extra_flags": "",
+   "cls_features": "ep",
+   "ep_queries": 32,
+   "d_out": 1,
+   "head_epoch": 89,
+   "val_acc1_at_head_epoch": 81.32,
+   "table_best_acc1": 81.39,
+   "note": "final-epoch head, not best-epoch: checkpoints overwrite per epoch. table_best_acc1 is the peak of the same run; the difference is the post-peak decline of a frozen-feature probe, largest on VLM encoders.",
+   "training_log": "logs/beitv2_vitb/ep.txt",
+   "params_incl_bn_stats": 1384937,
+   "file": "beitv2_vitb/ep_head.pth",
+   "size_mb": 5.5
+  },
+  {
+   "method": "SimMIM",
+   "arch": "ViT-B/16",
+   "pretrain": "IN-1K",
+   "image_size": 224,
+   "model": "vit_base_patch16",
+   "loader": "simmim",
+   "pretrain_tag": "",
+   "finetune": "vit_base_patch16_224.simmim/checkpoint-799.pth",
+   "extra_flags": "",
+   "cls_features": "ep",
+   "ep_queries": 32,
+   "d_out": 1,
+   "head_epoch": 89,
+   "val_acc1_at_head_epoch": 64.81,
+   "table_best_acc1": 64.86,
+   "note": "final-epoch head, not best-epoch: checkpoints overwrite per epoch. table_best_acc1 is the peak of the same run; the difference is the post-peak decline of a frozen-feature probe, largest on VLM encoders.",
+   "training_log": "logs/simmim_vitb/ep.txt",
+   "params_incl_bn_stats": 1384937,
+   "file": "simmim_vitb/ep_head.pth",
+   "size_mb": 5.5
+  },
+  {
+   "method": "CAPI",
+   "arch": "ViT-L/14",
+   "pretrain": "IN-1K",
+   "image_size": 224,
+   "model": "capi_vitl14_in1k",
+   "loader": "capi",
+   "pretrain_tag": "",
+   "finetune": "",
+   "extra_flags": "",
+   "cls_features": "ep",
+   "ep_queries": 32,
+   "d_out": 1,
+   "head_epoch": 89,
+   "val_acc1_at_head_epoch": 82.43,
+   "table_best_acc1": 82.98,
+   "note": "final-epoch head, not best-epoch: checkpoints overwrite per epoch. table_best_acc1 is the peak of the same run; the difference is the post-peak decline of a frozen-feature probe, largest on VLM encoders.",
+   "training_log": "logs/capi_vitl/ep.txt",
+   "params_incl_bn_stats": 2108393,
+   "file": "capi_vitl/ep_head.pth",
+   "size_mb": 8.4
+  },
+  {
+   "method": "DINO",
+   "arch": "ViT-B/16",
+   "pretrain": "IN-1K",
+   "image_size": 224,
+   "model": "vit_base_patch16",
+   "loader": "",
+   "pretrain_tag": "",
+   "finetune": "vit_base_patch16_224.dino",
+   "extra_flags": "",
+   "cls_features": "ep_all",
+   "ep_queries": 32,
+   "d_out": 1,
+   "head_epoch": 89,
+   "val_acc1_at_head_epoch": 77.08,
+   "table_best_acc1": 77.39,
+   "note": "final-epoch head, not best-epoch: checkpoints overwrite per epoch. table_best_acc1 is the peak of the same run; the difference is the post-peak decline of a frozen-feature probe, largest on VLM encoders.",
+   "training_log": "logs/dino_vitb/ep.txt",
+   "params_incl_bn_stats": 1384937,
+   "file": "dino_vitb/ep_head.pth",
+   "size_mb": 5.5
+  },
+  {
+   "method": "iBOT",
+   "arch": "ViT-B/16",
+   "pretrain": "IN-1K",
+   "image_size": 224,
+   "model": "vit_base_patch16",
+   "loader": "",
+   "pretrain_tag": "",
+   "finetune": "vit_base_patch16_224.ibot/checkpoint-799.pth",
+   "extra_flags": "",
+   "cls_features": "ep_all",
+   "ep_queries": 32,
+   "d_out": 1,
+   "head_epoch": 89,
+   "val_acc1_at_head_epoch": 78.62,
+   "table_best_acc1": 78.98,
+   "note": "final-epoch head, not best-epoch: checkpoints overwrite per epoch. table_best_acc1 is the peak of the same run; the difference is the post-peak decline of a frozen-feature probe, largest on VLM encoders.",
+   "training_log": "logs/ibot_vitb/ep.txt",
+   "params_incl_bn_stats": 1384937,
+   "file": "ibot_vitb/ep_head.pth",
+   "size_mb": 5.5
+  },
+  {
+   "method": "DINOv2",
+   "arch": "ViT-B/14",
+   "pretrain": "LVD-142M",
+   "image_size": 224,
+   "model": "dinov2_vitb14",
+   "loader": "hub",
+   "pretrain_tag": "",
+   "finetune": "",
+   "extra_flags": "",
+   "cls_features": "ep",
+   "ep_queries": 32,
+   "d_out": 1,
+   "head_epoch": 25,
+   "val_acc1_at_head_epoch": 83.61,
+   "table_best_acc1": 84.01,
+   "note": "final-epoch head, not best-epoch: checkpoints overwrite per epoch. table_best_acc1 is the peak of the same run; the difference is the post-peak decline of a frozen-feature probe, largest on VLM encoders.",
+   "training_log": "logs/dinov2_vitb/ep.txt",
+   "params_incl_bn_stats": 1384937,
+   "file": "dinov2_vitb/ep_head.pth",
+   "size_mb": 5.5
+  },
+  {
+   "method": "DINOv2",
+   "arch": "ViT-L/14",
+   "pretrain": "LVD-142M",
+   "image_size": 224,
+   "model": "dinov2_vitl14",
+   "loader": "hub",
+   "pretrain_tag": "",
+   "finetune": "",
+   "extra_flags": "",
+   "cls_features": "ep_all",
+   "ep_queries": 32,
+   "d_out": 1,
+   "head_epoch": 15,
+   "val_acc1_at_head_epoch": 85.56,
+   "table_best_acc1": 85.56,
+   "note": "best-epoch head: saved at the run's peak, the number the table reports.",
+   "training_log": "logs/dinov2_vitl/ep.txt",
+   "params_incl_bn_stats": 2108393,
+   "file": "dinov2_vitl/ep_head.pth",
+   "size_mb": 8.4
+  },
+  {
+   "method": "DINOv3",
+   "arch": "ViT-B/16",
+   "pretrain": "LVD-1689M",
+   "image_size": 224,
+   "model": "dinov3_vitb16",
+   "loader": "dinov3",
+   "pretrain_tag": "",
+   "finetune": "",
+   "extra_flags": "",
+   "cls_features": "ep_all",
+   "ep_queries": 32,
+   "d_out": 1,
+   "head_epoch": 20,
+   "val_acc1_at_head_epoch": 83.77,
+   "table_best_acc1": 84.11,
+   "note": "final-epoch head, not best-epoch: checkpoints overwrite per epoch. table_best_acc1 is the peak of the same run; the difference is the post-peak decline of a frozen-feature probe, largest on VLM encoders.",
+   "training_log": "logs/dinov3_vitb/ep.txt",
+   "params_incl_bn_stats": 1384937,
+   "file": "dinov3_vitb/ep_head.pth",
+   "size_mb": 5.5
+  },
+  {
+   "method": "DINOv3",
+   "arch": "ViT-L/16",
+   "pretrain": "LVD-1689M",
+   "image_size": 224,
+   "model": "dinov3_vitl16",
+   "loader": "dinov3",
+   "pretrain_tag": "",
+   "finetune": "",
+   "extra_flags": "",
+   "cls_features": "ep_all",
+   "ep_queries": 32,
+   "d_out": 1,
+   "head_epoch": 19,
+   "val_acc1_at_head_epoch": 86.73,
+   "table_best_acc1": 87.05,
+   "note": "final-epoch head, not best-epoch: checkpoints overwrite per epoch. table_best_acc1 is the peak of the same run; the difference is the post-peak decline of a frozen-feature probe, largest on VLM encoders.",
+   "training_log": "logs/dinov3_vitl/ep.txt",
+   "params_incl_bn_stats": 2108393,
+   "file": "dinov3_vitl/ep_head.pth",
+   "size_mb": 8.4
+  },
+  {
+   "method": "CLIP",
+   "arch": "ViT-L/14",
+   "pretrain": "WIT-400M",
+   "image_size": 224,
+   "model": "ViT-L-14",
+   "loader": "openclip",
+   "pretrain_tag": "openai",
+   "finetune": "",
+   "extra_flags": "",
+   "cls_features": "ep",
+   "ep_queries": 32,
+   "d_out": 1,
+   "head_epoch": 11,
+   "val_acc1_at_head_epoch": 83.22,
+   "table_best_acc1": 83.22,
+   "note": "best-epoch head: saved at the run's peak, the number the table reports.",
+   "training_log": "logs/clip_vitl/ep.txt",
+   "params_incl_bn_stats": 2108393,
+   "file": "clip_vitl/ep_head.pth",
+   "size_mb": 8.4
+  },
+  {
+   "method": "SigLIP",
+   "arch": "ViT-L/16",
+   "pretrain": "WebLI (10B imgs, 109 langs)",
+   "image_size": 256,
+   "model": "ViT-L-16-SigLIP-256",
+   "loader": "openclip",
+   "pretrain_tag": "webli",
+   "finetune": "",
+   "extra_flags": "",
+   "cls_features": "ep",
+   "ep_queries": 32,
+   "d_out": 1,
+   "head_epoch": 6,
+   "val_acc1_at_head_epoch": 85.93,
+   "table_best_acc1": 85.93,
+   "note": "best-epoch head: saved at the run's peak, the number the table reports.",
+   "training_log": "logs/siglip_vitl/ep.txt",
+   "params_incl_bn_stats": 2108393,
+   "file": "siglip_vitl/ep_head.pth",
+   "size_mb": 8.4
+  },
+  {
+   "method": "SigLIP2",
+   "arch": "ViT-L/16",
+   "pretrain": "WebLI (10B imgs, 109 langs)",
+   "image_size": 256,
+   "model": "ViT-L-16-SigLIP2-256",
+   "loader": "openclip",
+   "pretrain_tag": "webli",
+   "finetune": "",
+   "extra_flags": "",
+   "cls_features": "ep",
+   "ep_queries": 32,
+   "d_out": 1,
+   "head_epoch": 6,
+   "val_acc1_at_head_epoch": 87.06,
+   "table_best_acc1": 87.06,
+   "note": "best-epoch head: saved at the run's peak, the number the table reports.",
+   "training_log": "logs/siglip2_vitl/ep.txt",
+   "params_incl_bn_stats": 2108393,
+   "file": "siglip2_vitl/ep_head.pth",
+   "size_mb": 8.4
+  },
+  {
+   "method": "AIMv2",
+   "arch": "ViT-L/14",
+   "pretrain": "DFN-2B + COYO + HQITP",
+   "image_size": 224,
+   "model": "aimv2-large-patch14-224",
+   "loader": "aimv2",
+   "pretrain_tag": "",
+   "finetune": "",
+   "extra_flags": "",
+   "cls_features": "ep",
+   "ep_queries": 32,
+   "d_out": 1,
+   "head_epoch": 19,
+   "val_acc1_at_head_epoch": 85.62,
+   "table_best_acc1": 85.94,
+   "note": "final-epoch head, not best-epoch: checkpoints overwrite per epoch. table_best_acc1 is the peak of the same run; the difference is the post-peak decline of a frozen-feature probe, largest on VLM encoders.",
+   "training_log": "logs/aimv2_vitl/ep.txt",
+   "params_incl_bn_stats": 2108393,
+   "file": "aimv2_vitl/ep_head.pth",
+   "size_mb": 8.4
+  },
+  {
+   "method": "SigLIP2",
+   "arch": "SO400M/14",
+   "pretrain": "WebLI (10B imgs, 109 langs)",
+   "image_size": 224,
+   "model": "ViT-SO400M-14-SigLIP2",
+   "loader": "openclip",
+   "pretrain_tag": "webli",
+   "finetune": "",
+   "extra_flags": "",
+   "cls_features": "ep",
+   "ep_queries": 32,
+   "d_out": 1,
+   "head_epoch": 29,
+   "val_acc1_at_head_epoch": 85.64,
+   "table_best_acc1": 87.68,
+   "note": "final-epoch head, not best-epoch: checkpoints overwrite per epoch. table_best_acc1 is the peak of the same run; the difference is the post-peak decline of a frozen-feature probe, largest on VLM encoders.",
+   "training_log": "logs/siglip2_so400m/ep.txt",
+   "params_incl_bn_stats": 2519273,
+   "file": "siglip2_so400m/ep_head.pth",
+   "size_mb": 10.1
+  },
+  {
+   "method": "EVA02",
+   "arch": "ViT-L/14",
+   "pretrain": "IN-22K",
+   "image_size": 224,
+   "model": "eva02_large_patch14_224.mim_in22k",
+   "loader": "timm",
+   "pretrain_tag": "",
+   "finetune": "",
+   "extra_flags": "",
+   "cls_features": "ep",
+   "ep_queries": 32,
+   "d_out": 1,
+   "head_epoch": 89,
+   "val_acc1_at_head_epoch": 83.22,
+   "table_best_acc1": 83.58,
+   "note": "final-epoch head, not best-epoch: checkpoints overwrite per epoch. table_best_acc1 is the peak of the same run; the difference is the post-peak decline of a frozen-feature probe, largest on VLM encoders.",
+   "training_log": "logs/eva02_vitl_mim/ep.txt",
+   "params_incl_bn_stats": 2108393,
+   "file": "eva02_vitl_mim/ep_head.pth",
+   "size_mb": 8.4
+  },
+  {
+   "method": "RADIO",
+   "arch": "ViT-L/16",
+   "pretrain": "distilled (DFN5B/SigLIP/DINOv2/SAM)",
+   "image_size": 224,
+   "model": "radio_v2.5-l",
+   "loader": "radio",
+   "pretrain_tag": "",
+   "finetune": "",
+   "extra_flags": "",
+   "cls_features": "ep",
+   "ep_queries": 32,
+   "d_out": 1,
+   "head_epoch": 89,
+   "val_acc1_at_head_epoch": 83.4,
+   "table_best_acc1": 83.89,
+   "note": "final-epoch head, not best-epoch: checkpoints overwrite per epoch. table_best_acc1 is the peak of the same run; the difference is the post-peak decline of a frozen-feature probe, largest on VLM encoders.",
+   "training_log": "logs/radio_v25_l/ep.txt",
+   "params_incl_bn_stats": 2108393,
+   "file": "radio_v25_l/ep_head.pth",
+   "size_mb": 8.4
+  },
+  {
+   "method": "Hiera",
+   "arch": "ViT-H/16",
+   "pretrain": "IN-1K",
+   "image_size": 224,
+   "model": "hiera_huge_224.mae",
+   "loader": "timm",
+   "pretrain_tag": "",
+   "finetune": "",
+   "extra_flags": "",
+   "cls_features": "ep",
+   "ep_queries": 32,
+   "d_out": 1,
+   "head_epoch": 89,
+   "val_acc1_at_head_epoch": 79.82,
+   "table_best_acc1": 79.89,
+   "note": "final-epoch head, not best-epoch: checkpoints overwrite per epoch. table_best_acc1 is the peak of the same run; the difference is the post-peak decline of a frozen-feature probe, largest on VLM encoders.",
+   "training_log": "logs/hiera_huge/ep.txt",
+   "params_incl_bn_stats": 6312937,
+   "file": "hiera_huge/ep_head.pth",
+   "size_mb": 25.3
+  },
+  {
+   "method": "Hiera",
+   "arch": "ViT-L/16",
+   "pretrain": "IN-1K",
+   "image_size": 224,
+   "model": "hiera_large_224.mae",
+   "loader": "timm",
+   "pretrain_tag": "",
+   "finetune": "",
+   "extra_flags": "",
+   "cls_features": "ep",
+   "ep_queries": 32,
+   "d_out": 1,
+   "head_epoch": 83,
+   "val_acc1_at_head_epoch": 78.51,
+   "table_best_acc1": 78.55,
+   "note": "final-epoch head, not best-epoch: checkpoints overwrite per epoch. table_best_acc1 is the peak of the same run; the difference is the post-peak decline of a frozen-feature probe, largest on VLM encoders.",
+   "training_log": "logs/hiera_large/ep.txt",
+   "params_incl_bn_stats": 2519273,
+   "file": "hiera_large/ep_head.pth",
+   "size_mb": 10.1
+  },
+  {
+   "method": "Hiera",
+   "arch": "ViT-B/16",
+   "pretrain": "IN-1K",
+   "image_size": 224,
+   "model": "hiera_base_224.mae",
+   "loader": "timm",
+   "pretrain_tag": "",
+   "finetune": "",
+   "extra_flags": "",
+   "cls_features": "ep",
+   "ep_queries": 32,
+   "d_out": 1,
+   "head_epoch": 88,
+   "val_acc1_at_head_epoch": 75.63,
+   "table_best_acc1": 75.68,
+   "note": "final-epoch head, not best-epoch: checkpoints overwrite per epoch. table_best_acc1 is the peak of the same run; the difference is the post-peak decline of a frozen-feature probe, largest on VLM encoders.",
+   "training_log": "logs/hiera_base/ep.txt",
+   "params_incl_bn_stats": 1384937,
+   "file": "hiera_base/ep_head.pth",
+   "size_mb": 5.5
+  },
+  {
+   "method": "RADIO",
+   "arch": "ViT-B/16",
+   "pretrain": "distilled (DFN5B/SigLIP/DINOv2/SAM)",
+   "image_size": 224,
+   "model": "radio_v2.5-b",
+   "loader": "radio",
+   "pretrain_tag": "",
+   "finetune": "",
+   "extra_flags": "",
+   "cls_features": "ep",
+   "ep_queries": 32,
+   "d_out": 1,
+   "head_epoch": 89,
+   "val_acc1_at_head_epoch": 80.26,
+   "table_best_acc1": 80.29,
+   "note": "final-epoch head, not best-epoch: checkpoints overwrite per epoch. table_best_acc1 is the peak of the same run; the difference is the post-peak decline of a frozen-feature probe, largest on VLM encoders.",
+   "training_log": "logs/radio_v25_b/ep.txt",
+   "params_incl_bn_stats": 1384937,
+   "file": "radio_v25_b/ep_head.pth",
+   "size_mb": 5.5
+  },
+  {
+   "method": "CLIP",
+   "arch": "ViT-B/16",
+   "pretrain": "WIT-400M",
+   "image_size": 224,
+   "model": "ViT-B-16",
+   "loader": "openclip",
+   "pretrain_tag": "openai",
+   "finetune": "",
+   "extra_flags": "",
+   "cls_features": "ep_all",
+   "ep_queries": 32,
+   "d_out": 1,
+   "head_epoch": 11,
+   "val_acc1_at_head_epoch": 77.85,
+   "table_best_acc1": 77.85,
+   "note": "best-epoch head: saved at the run's peak, the number the table reports.",
+   "training_log": "logs/clip_vitb16/ep.txt",
+   "params_incl_bn_stats": 792553,
+   "file": "clip_vitb16/ep_head.pth",
+   "size_mb": 3.2
+  },
+  {
+   "method": "PE-Core",
+   "arch": "L-14/336",
+   "pretrain": "MetaCLIP-curated 5.4B",
+   "image_size": 336,
+   "model": "PE-Core-L-14-336",
+   "loader": "openclip",
+   "pretrain_tag": "meta",
+   "finetune": "",
+   "extra_flags": "",
+   "cls_features": "ep",
+   "ep_queries": 32,
+   "d_out": 1,
+   "head_epoch": 12,
+   "val_acc1_at_head_epoch": 87.25,
+   "table_best_acc1": 87.25,
+   "note": "best-epoch head: saved at the run's peak, the number the table reports.",
+   "training_log": "logs/pe_core_l336/ep.txt",
+   "params_incl_bn_stats": 2108393,
+   "file": "pe_core_l336/ep_head.pth",
+   "size_mb": 8.4
+  },
+  {
+   "method": "EVA02-CLIP",
+   "arch": "E-14",
+   "pretrain": "LAION-2B (4B seen)",
+   "image_size": 224,
+   "model": "EVA02-E-14",
+   "loader": "openclip",
+   "pretrain_tag": "laion2b_s4b_b115k",
+   "finetune": "",
+   "extra_flags": "",
+   "cls_features": "ep",
+   "ep_queries": 32,
+   "d_out": 1,
+   "head_epoch": 6,
+   "val_acc1_at_head_epoch": 87.7,
+   "table_best_acc1": 87.7,
+   "note": "best-epoch head: saved at the run's peak, the number the table reports.",
+   "training_log": "logs/eva02_e14/ep.txt",
+   "params_incl_bn_stats": 5065193,
+   "file": "eva02_e14/ep_head.pth",
+   "size_mb": 20.3
+  },
+  {
+   "method": "MetaCLIP2",
+   "arch": "ViT-bigG/14",
+   "pretrain": "MetaCLIP2-worldwide (300+ langs)",
+   "image_size": 224,
+   "model": "ViT-bigG-14-worldwide",
+   "loader": "openclip",
+   "pretrain_tag": "metaclip2_worldwide",
+   "finetune": "",
+   "extra_flags": "",
+   "cls_features": "ep",
+   "ep_queries": 32,
+   "d_out": 1,
+   "head_epoch": 6,
+   "val_acc1_at_head_epoch": 87.11,
+   "table_best_acc1": 87.11,
+   "note": "best-epoch head: saved at the run's peak, the number the table reports.",
+   "training_log": "logs/metaclip2_bigg/ep.txt",
+   "params_incl_bn_stats": 4490473,
+   "file": "metaclip2_bigg/ep_head.pth",
+   "size_mb": 18.0
+  },
+  {
+   "method": "DINOv3",
+   "arch": "ViT-7B/16",
+   "pretrain": "LVD-1689M",
+   "image_size": 224,
+   "model": "dinov3_vit7b16",
+   "loader": "dinov3",
+   "pretrain_tag": "",
+   "finetune": "",
+   "extra_flags": "",
+   "cls_features": "ep_all",
+   "ep_queries": 32,
+   "d_out": 1,
+   "head_epoch": 6,
+   "val_acc1_at_head_epoch": 88.36,
+   "table_best_acc1": 88.36,
+   "note": "best-epoch head: saved at the run's peak, the number the table reports.",
+   "training_log": "logs/dinov3_vit7b/ep.txt",
+   "params_incl_bn_stats": 21013481,
+   "file": "dinov3_vit7b/ep_head.pth",
+   "size_mb": 84.1
+  },
+  {
+   "method": "MaskFeat",
+   "arch": "ViT-B/16",
+   "pretrain": "IN-1K",
+   "image_size": 224,
+   "model": "vit_base_patch16",
+   "loader": "",
+   "pretrain_tag": "",
+   "finetune": "vit_base_patch16_224.maskfeat/checkpoint-799.pth",
+   "extra_flags": "",
+   "cls_features": "ep",
+   "ep_queries": 32,
+   "d_out": 1,
+   "head_epoch": 89,
+   "val_acc1_at_head_epoch": 71.68,
+   "table_best_acc1": 71.76,
+   "note": "final-epoch head, not best-epoch: checkpoints overwrite per epoch. table_best_acc1 is the peak of the same run; the difference is the post-peak decline of a frozen-feature probe, largest on VLM encoders.",
+   "training_log": "logs/maskfeat_vitb/ep.txt",
+   "params_incl_bn_stats": 1384937,
+   "file": "maskfeat_vitb/ep_head.pth",
+   "size_mb": 5.5
+  },
+  {
+   "method": "I-JEPA",
+   "arch": "ViT-H/14",
+   "pretrain": "IN-1K",
+   "image_size": 224,
+   "model": "vit_huge_patch14",
+   "loader": "",
+   "pretrain_tag": "",
+   "finetune": "vit_huge_patch14_224.ijepa/checkpoint-799.pth",
+   "extra_flags": "--no_cls_token",
+   "cls_features": "ep",
+   "ep_queries": 32,
+   "d_out": 1,
+   "head_epoch": 89,
+   "val_acc1_at_head_epoch": 78.8,
+   "table_best_acc1": 79.0,
+   "note": "final-epoch head, not best-epoch: checkpoints overwrite per epoch. table_best_acc1 is the peak of the same run; the difference is the post-peak decline of a frozen-feature probe, largest on VLM encoders.",
+   "training_log": "logs/ijepa_vith/ep.txt",
+   "params_incl_bn_stats": 2962921,
+   "file": "ijepa_vith/ep_head.pth",
+   "size_mb": 11.9
+  },
+  {
+   "method": "MoCov3",
+   "arch": "ViT-B/16",
+   "pretrain": "IN-1K",
+   "image_size": 224,
+   "model": "vit_base_patch16",
+   "loader": "",
+   "pretrain_tag": "",
+   "finetune": "vit_base_patch16_224.mocov3/checkpoint-799.pth",
+   "extra_flags": "",
+   "cls_features": "ep_all",
+   "ep_queries": 32,
+   "d_out": 1,
+   "head_epoch": 89,
+   "val_acc1_at_head_epoch": 76.21,
+   "table_best_acc1": 76.53,
+   "note": "final-epoch head, not best-epoch: checkpoints overwrite per epoch. table_best_acc1 is the peak of the same run; the difference is the post-peak decline of a frozen-feature probe, largest on VLM encoders.",
+   "training_log": "logs/mocov3_vitb/ep.txt",
+   "params_incl_bn_stats": 1384937,
+   "file": "mocov3_vitb/ep_head.pth",
+   "size_mb": 5.5
+  },
+  {
+   "method": "iBOT",
+   "arch": "ViT-L/16",
+   "pretrain": "IN-1K",
+   "image_size": 224,
+   "model": "vit_large_patch16",
+   "loader": "",
+   "pretrain_tag": "",
+   "finetune": "vit_large_patch16_224.ibot/checkpoint-799.pth",
+   "extra_flags": "",
+   "cls_features": "ep_all",
+   "ep_queries": 32,
+   "d_out": 1,
+   "head_epoch": 89,
+   "val_acc1_at_head_epoch": 79.43,
+   "table_best_acc1": 79.97,
+   "note": "final-epoch head, not best-epoch: checkpoints overwrite per epoch. table_best_acc1 is the peak of the same run; the difference is the post-peak decline of a frozen-feature probe, largest on VLM encoders.",
+   "training_log": "logs/ibot_vitl/ep.txt",
+   "params_incl_bn_stats": 2108393,
+   "file": "ibot_vitl/ep_head.pth",
+   "size_mb": 8.4
+  },
+  {
+   "method": "MetaCLIP2",
+   "arch": "ViT-bigG/14-378",
+   "pretrain": "MetaCLIP2-worldwide (300+ langs)",
+   "image_size": 378,
+   "model": "ViT-bigG-14-worldwide-378",
+   "loader": "openclip",
+   "pretrain_tag": "metaclip2_worldwide",
+   "finetune": "",
+   "extra_flags": "",
+   "cls_features": "ep",
+   "ep_queries": 32,
+   "d_out": 1,
+   "head_epoch": 6,
+   "val_acc1_at_head_epoch": 88.12,
+   "table_best_acc1": 88.12,
+   "note": "best-epoch head: saved at the run's peak, the number the table reports.",
+   "training_log": "logs/metaclip2_bigg378/ep.txt",
+   "params_incl_bn_stats": 4490473,
+   "file": "metaclip2_bigg378/ep_head.pth",
+   "size_mb": 18.0
+  },
+  {
+   "method": "MaskFeat",
+   "arch": "ViT-L/16",
+   "pretrain": "IN-1K",
+   "image_size": 224,
+   "model": "vit_large_patch16",
+   "loader": "",
+   "pretrain_tag": "",
+   "finetune": "vit_large_patch16_224.maskfeat/checkpoint-799.pth",
+   "extra_flags": "",
+   "cls_features": "ep",
+   "ep_queries": 32,
+   "d_out": 1,
+   "head_epoch": 89,
+   "val_acc1_at_head_epoch": 69.56,
+   "table_best_acc1": 69.63,
+   "note": "final-epoch head, not best-epoch: checkpoints overwrite per epoch. table_best_acc1 is the peak of the same run; the difference is the post-peak decline of a frozen-feature probe, largest on VLM encoders.",
+   "training_log": "logs/maskfeat_large/ep.txt",
+   "params_incl_bn_stats": 2108393,
+   "file": "maskfeat_large/ep_head.pth",
+   "size_mb": 8.4
+  },
+  {
+   "method": "EVA02-CLIP",
+   "arch": "E-14-plus",
+   "pretrain": "LAION-2B (9B seen)",
+   "image_size": 224,
+   "model": "EVA02-E-14-plus",
+   "loader": "openclip",
+   "pretrain_tag": "laion2b_s9b_b144k",
+   "finetune": "",
+   "extra_flags": "",
+   "cls_features": "ep",
+   "ep_queries": 32,
+   "d_out": 1,
+   "head_epoch": 6,
+   "val_acc1_at_head_epoch": 87.98,
+   "table_best_acc1": 87.98,
+   "note": "best-epoch head: saved at the run's peak, the number the table reports.",
+   "training_log": "logs/eva02_e14_plus/ep.txt",
+   "params_incl_bn_stats": 5065193,
+   "file": "eva02_e14_plus/ep_head.pth",
+   "size_mb": 20.3
+  },
+  {
+   "method": "Franca",
+   "arch": "ViT-L/14",
+   "pretrain": "LAION-600M",
+   "image_size": 224,
+   "model": "franca_vitl14",
+   "loader": "franca",
+   "pretrain_tag": "",
+   "finetune": "",
+   "extra_flags": "--franca_weights LAION",
+   "cls_features": "ep_all",
+   "ep_queries": 32,
+   "d_out": 1,
+   "head_epoch": 14,
+   "val_acc1_at_head_epoch": 84.28,
+   "table_best_acc1": 84.28,
+   "note": "best-epoch head: saved at the run's peak, the number the table reports.",
+   "training_log": "logs/franca_laion/ep_all.txt",
+   "params_incl_bn_stats": 2108393,
+   "file": "franca_laion/ep_head.pth",
+   "size_mb": 8.4
+  }
+ ],
+ "missing": [
+  {
+   "method": "DiT",
+   "arch": "DiT-XL/2",
+   "reason": "no surviving checkpoint: the run's output directory was deleted after its logs were published; re-run required to export a head"
+  }
+ ]
+}