MIT-OL-AI-D/cv-nlp-lab5

CV & NLP Lab 5: unpaired alignment embeddings

Teaching data for Lab 5 of the MIT CV & NLP MicroMasters course (6.4600.5x), "Aligning Two Models That Never Met". Learners map image embeddings from DINOv2 into caption embeddings from MPNet without any image-caption pairs, and score the map on held-out pairs and on zero-shot CIFAR-10 classification. The bundle also holds the paired maps the lab compares against; learners fit every map without pairs themselves.

The files hold embeddings, caption text, object kinds and paired reference maps. They contain no images.

Files

File	Holds
lab5_embeddings.npz	DINOv2 ViT-B/14 image embeddings and MPNet caption embeddings, float16, 768 dimensions, as named arrays. Load with numpy.load(path, allow_pickle=False)
lab5_maps.npz	three paired reference maps, float32, as named arrays. Load the same way
lab5_text.json	for each of the 4,000 held-out images: its COCO image ID, every caption, and its kind; plus the CIFAR-10 class names

Arrays in lab5_embeddings.npz	Rows	Use
images, captions	20,000 each, from different COCO train images	fitting maps with no pairs
pair_images, pair_captions	2,000, row-aligned	the paired reference maps only
animals_images, animals_captions, indoor_images, indoor_captions	6,000 images and the captions of 6,000 other images, per kind	the kind experiment
val_images, val_captions	4,000 COCO val images, row-aligned	scoring only
val_animals_*, val_indoor_*	1,500 per kind, row-aligned	scoring the kind experiment
cifar_images, cifar_labels, cifar_prompts	5,000 images, 10 class prompts	zero-shot scoring

Arrays in lab5_maps.npz	Shape	What it is
pairs_50, pairs_100, pairs_200	768 x 768	orthogonal Procrustes maps fitted on the first 50, 100 and 200 rows of the true pairs

A map takes an image embedding x, centered on the unpaired images' mean and scaled to unit length, to xW in caption space.

Sources and licenses

• Embeddings: a subset of schnaus/unpaired-rosetta-embeddings, released with Schnaus et al. (2026). That dataset is licensed "source-dataset-licenses": each source below keeps its own terms.
• MS COCO 2014: the annotations, captions and object labels both, are CC BY 4.0. The images are under Flickr's terms of use and are not included; the image IDs link to cocodataset.org.
• CIFAR-10: Krizhevsky (2009), distributed without a stated license.
• DINOv2 (Meta AI) and all-mpnet-base-v2 (sentence-transformers): Apache 2.0.

How it was built

The subset is drawn with seed 6465 from the embeddings above. COCO 2014 rows are in image-ID order. An image's captions are averaged after each is normalized. CIFAR-10 images are drawn at random. An image's kind comes from its COCO object labels: animals if it shows an animal and nothing in the indoor, kitchen, food, furniture, appliance or electronic groups, indoor for the reverse, and other for everything else. The maps were fitted on an NVIDIA T4 by orthogonal Procrustes on true pairs.

Citations

@article{schnaus2026rosetta,
  title   = {Shared Geometry as a Rosetta Stone: Cross-Modal Alignment Without Paired Data},
  author  = {Schnaus, Dominik and Dag{\`e}s, Thomas and Cremers, Daniel and Wang, Xi and Isola, Phillip},
  journal = {arXiv preprint arXiv:2610.09411},
  year    = {2026}
}
@inproceedings{lin2014coco,
  title     = {Microsoft {COCO}: Common Objects in Context},
  author    = {Lin, Tsung-Yi and others},
  booktitle = {European Conference on Computer Vision (ECCV)},
  year      = {2014}
}
@techreport{krizhevsky2009cifar,
  title       = {Learning Multiple Layers of Features from Tiny Images},
  author      = {Krizhevsky, Alex},
  institution = {University of Toronto},
  year        = {2009}
}
@article{oquab2024dinov2,
  title   = {{DINOv2}: Learning Robust Visual Features without Supervision},
  author  = {Oquab, Maxime and others},
  journal = {Transactions on Machine Learning Research},
  year    = {2024}
}
@inproceedings{song2020mpnet,
  title     = {{MPNet}: Masked and Permuted Pre-training for Language Understanding},
  author    = {Song, Kaitao and Tan, Xu and Qin, Tao and Lu, Jianfeng and Liu, Tie-Yan},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
  year      = {2020}
}