docs: add training-data scale (~0.47B samples, 96% image / 4% video) + data-efficiency vs DINOv3 LVD-1689M

README.md

@@ -79,6 +79,7 @@ with torch.inference_mode():
 | Video tokenizer | 3D Conv, tubelet size 2 |
 | Precision | bf16 weights |
 | Training | DINO + iBOT + KoLeo self-distillation, Gram anchoring, contrastive caption alignment |
+| Training data | ~0.47B samples — 448.6M images (96%) + 18.5M video clips (4%) |
 
 Outputs (`BaseModelOutputWithPooling`): `last_hidden_state` `(B, 1+4+N, 4096)`,
 `pooler_output` `(B, 4096)`.
@@ -101,8 +102,12 @@ better for every column **except KITTI depth MSE** (lower is better). Best value
   reflecting the encoder's dense, temporally-coherent patch features on video.
 - **ImageNet-1K linear probe (87.2)** and **ADE20K semantic segmentation (52.0 mIoU)** are
   second only to DINOv3 7B while ahead of DINOv2, V-JEPA 2.1, and SigLIP2.
+- **Data efficiency** — these results come from **~0.47B training samples** (448.6M images +
+  18.5M video clips), roughly **3.6× less data than DINOv3**, which is trained on LVD-1689M
+  (1,689M images). Despite the smaller corpus — and with only ~4% of samples being video — the
+  encoder stays competitive across both image and video benchmarks, and leads on DAVIS.
 - K400 action recognition and KITTI depth estimation are still being evaluated and will be added
-  as the runs complete.
+  as the runs complete. Numbers in this table will be updated as evaluations finish.
 
 Protocol: DINOv3-style linear/attentive probes for image tasks; V-JEPA 2-style protocol for
 video. Comparison numbers are the best figures reported by each model's authors.