Meldashti/vlm-compositionality-embeddings

VLM Compositionality Embeddings

Pre-computed image and text embeddings for the thesis "From Euclidean to Hyperbolic Vision-Language Spaces: A Study of Attribute–Object Compositionality" by Meelad Dashti (Politecnico di Torino & University of Twente, 2026).

Code repository: github.com/MelDashti/hyperbolic-vlm-compositionality

Models

Model	Geometry	Architecture	Training Data
CLIP ViT-L/14	Spherical	ViT-L/14	WIT (400M+ pairs)
DINOv2 ViT-L/14	Spherical	ViT-L/14	LVD-142M (self-supervised)
CLIP-B (GRIT)	Spherical	ViT-B/16	GRIT (20.5M pairs)
MERU-B (GRIT)	Hyperbolic	ViT-B/16	GRIT (20.5M pairs)
HyCoCLIP-B (GRIT)	Hyperbolic	ViT-B/16	GRIT (20.5M pairs)

Datasets

CZSL Benchmarks

• MIT-States — 53K images, 115 attributes, 245 objects
• UT-Zappos — 33K images, 16 attributes, 12 objects
• C-GQA — 39K images, 413 attributes, 674 objects
• VAW-CZSL — 92K images, 413 attributes, 541 objects

Group Robustness

• WaterBirds — Bird type classification (spurious: background)
• CelebA — Hair color classification (spurious: gender)

File Structure

Each dataset directory contains:

{dataset}/
├── IMGemb_{model}_{pretraining}.pt          # Image embeddings
├── TEXTemb_{model}_{pretraining}.pt         # Text pair embeddings
├── TEXTemb_primitives_{model}_{pretraining}.pt  # Primitive text embeddings
├── metadata_compositional-split-natural.t7  # Dataset metadata
└── compositional-split-natural/
    ├── train_pairs.txt
    ├── val_pairs.txt
    └── test_pairs.txt

File Naming Convention

• IMGemb_ — Image embeddings (one vector per image)
• TEXTemb_ — Text embeddings for (attribute, object) pair prompts
• TEXTemb_primitives_ — Separate attribute and object text embeddings
• Model identifiers: ViT-L-14_openai, CLIP-B_GRIT_GRIT, MERU-B_GRIT_GRIT, HyCoCLIP-B_HyCoCLIP, dinov2_vitl14_talk2dino, MERU-L_MERU

Usage

import torch

# Load image embeddings
img_emb = torch.load("mit-states/IMGemb_ViT-L-14_openai.pt", weights_only=False)

# Load text pair embeddings
text_emb = torch.load("mit-states/TEXTemb_ViT-L-14_openai.pt", weights_only=False)

# Load primitive text embeddings
primitives = torch.load("mit-states/TEXTemb_primitives_ViT-L-14_openai.pt", weights_only=False)
attr_embs = primitives['attr_embs']  # Individual attribute embeddings
obj_embs = primitives['obj_embs']    # Individual object embeddings

Download

# Clone with git LFS
git lfs install
git clone https://huggingface.co/datasets/Meldashti/vlm-compositionality-embeddings

# Or using huggingface_hub
from huggingface_hub import snapshot_download
snapshot_download("Meldashti/vlm-compositionality-embeddings", local_dir="data/", repo_type="dataset")

Citation

@mastersthesis{dashti2026euclidean,
  title={From Euclidean to Hyperbolic Vision-Language Spaces: A Study of Attribute-Object Compositionality},
  author={Dashti, Meelad},
  school={Politecnico di Torino \& University of Twente},
  year={2026}
}

License

MIT