region-embeddings

Per-region visual features for the Augustinian BabyLM project: every word-region pair from the grounding data, encoded with a frozen vision model.

Grounding comes from Flickr30k Entities, RefCOCO+, RefCOCOg and THINGS, together 563k region annotations. Each region is cropped and encoded; the region feature is the mean of the encoder's patch embeddings inside the bounding box. Three encoders are provided (DINOv3, iBOT ViT-B/16, SAM ViT-B), all trained on images alone with no text or caption supervision, so the features carry a purely perceptual prior.

Consumed by build_token_embeddings.py to produce augustinian-babylm/token-embeddings.

Part of https://github.com/bylinina/augustinian_babylm. Paper: https://openreview.net/forum?id=B4TD4XdlwF.

Citation

@inproceedings{bylinina2026augustinian,
  title     = {Augustinian BabyLM: What Ostensive Definition Can and Cannot
               Teach a Small Language Model},
  author    = {Bylinina, Lisa},
  booktitle = {Proceedings of the BabyLM Workshop},
  year      = {2026},
  url       = {https://openreview.net/forum?id=B4TD4XdlwF}
}