babyview-dino-vits16 — DINOv3 ViT-S trained from scratch on children's egocentric video

A 22M-parameter DINOv3 ViT-S/16 trained from scratch, self-supervised, on the BabyView corpus (Long et al., 2025): ~9.7M frames sampled at 1 fps from 2,701 hours of head-mounted-camera video of 51 children (ages 3–54 months) — release 2026.1. No internet images, no supervision, no pretrained initialization: this model has only ever seen a child's visual world.

Trained as part of "Scaling cross-situational word learning on children's naturalistic experiences" (Frank lab, Stanford; paper in prep) to test whether developmental-scale visual experience supports the representations needed for word learning.

Training

Stock DINOv3 recipe (github.com/awwkl/dinov3 fork of facebookresearch/dinov3), from scratch: global batch 512 (no gradient accumulation), 200k iterations (~102M samples seen), lr rule sqrt_wrt_1024, bf16, RoPE, Sinkhorn-Knopp centering + KoLeo + iBOT. Released weights are the EMA teacher backbone at iteration 199,999.

Evaluation (from the paper's probes)

• Konkle object prototype 4AFC (head-free object separability): high 90s — near internet- pretrained models.
• Cross-situational word-learning 4AFC after frozen-readout training on the full BabyView corpus: 39.6 (internet-pretrained DINOv3-S: high 70s–low 80s; chance 25). The persistent gap to internet pretraining — despite object separability — is the paper's central result about visual experience and word learning.

Two artifacts in this repo

1. HF-transformers weights (config.json + model.safetensors): load with AutoModel.from_pretrained. Note: the training run stored RoPE periods in bf16; the HF port recomputes them in fp32, so features differ microscopically from the training artifact (word-probe 29.1 ± 2.7 vs native 31.0 on our 100k probe — statistically indistinguishable).
2. vits_teacher_backbone_native.pt: the exact training artifact (native dinov3 state dict). Load with the dinov3 repo: build_model(cfg.student, only_teacher=True) then load_state_dict. The paper's numbers use this artifact.

Data statement

Model weights only. BabyView videos/frames are restricted human-subjects data and are NOT distributed here; corpus access is governed by the BabyView data-sharing process (Long et al., 2025, arXiv:2406.10447).

Acknowledgments

Training recipe and codebase groundwork by Khai Loong Aw (awwkl); BabyView corpus by the Frank lab and the BabyView team.