abdelstark/latent-inspector-fingerprints

latent-inspector fingerprints

Reference representation-geometry fingerprints for four self-supervised vision encoders — DINOv2 ViT-L/14, I-JEPA ViT-H/14, V-JEPA 2 ViT-L/16, and EUPE ViT-B/16 — computed on the same canonical image with latent-inspector.

This dataset is the numeric evidence layer behind the README table in abdelstark/vjepa2-vitl-fpc2-256-onnx. The ONNX exports in the Latent Inspector — ONNX Vision Encoders collection are the models; this dataset is what their patch embeddings look like when you actually probe them.

What's inside

File	Rows	Content
fingerprints.csv	4	Per-model patch-embedding geometry: effective rank, patch entropy, isotropy, spectral decay, RankMe, etc.
comparisons.csv	6	Pairwise cross-model alignment: linear CKA, k-NN overlap (k=10), patch correspondence
compare.json	1	The full latent-inspector compare report the CSVs were distilled from (raw metrics, pairwise matrices, overview highlights, metric caveats)

All four models are probed on a single shared input image: a canonical elephant photograph. The point of the dataset is not a large benchmark suite — it's a reproducible, auditable snapshot of what these four encoders do with identical pixels under identical normalization.

Why four models on one image?

Because the interesting question isn't "which model has the highest X" — it's how differently structured the representations are when the input is held constant.

• DINOv2 (ImageNet-centric self-distillation) — top-heavy, high CLS norm
• I-JEPA (predict-masked-latents, image-only) — high patch entropy, no CLS
• V-JEPA 2 (predict-masked-latents, video) — highest spatial coherence, highest CKA with DINOv2
• EUPE (proxy-distillation from a multi-expert teacher) — lowest effective rank, most top-heavy variance, no patch isotropy

All four are ViT-family image encoders. All four produce patch tokens. The same image goes in. The outputs have wildly different geometry. That's the point.

Schema

fingerprints.csv

Column	Meaning
model_name	Short identifier (dinov2-vit-l14, ijepa-vit-h14, vjepa2-vitl-img16-256, eupe-vit-b16)
n_patches	Number of spatial patch tokens in the model's output
embed_dim	Patch embedding dimensionality
effective_rank	Participation ratio of the patch covariance spectrum; rough measure of how many directions the model actually uses
dead_dimensions	Patch dimensions with zero variance
patch_entropy	Shannon entropy of the normalized singular-value distribution (higher = more spread)
cls_l2_norm	L2 norm of the CLS token (null for models without one)
patch_norm_mean / patch_norm_std	Mean and std of per-patch L2 norms
top10_variance_pct	Percentage of variance explained by the top 10 principal components
components_90pct	Number of PCs needed to explain 90% of variance
patch_isotropy	How isotropic the patch distribution is in the embedding space (0 = highly anisotropic, 1 = isotropic)
patch_uniformity	Log-avg pairwise Gaussian kernel on patch embeddings (higher = more uniform)
spatial_coherence	Average cosine similarity between spatially adjacent patches
rankme	RankMe estimator of representation rank
spectral_decay	Slope of the log-log covariance spectrum

comparisons.csv

Column	Meaning
model_a / model_b	The two models being compared
compared_patch_count	Number of shared patches used for comparison (truncated to the smaller grid when grids differ)
linear_cka	Linear CKA between the two patch matrices
knn_overlap_k10	Fraction of k=10 nearest-neighbor overlap across patches
cls_cosine_sim	CLS cosine similarity (null when CLS tokens are unavailable or dimension-incompatible)
mean_patch_correspondence	Mean per-patch cosine similarity after Procrustes alignment (null when embedding dims differ)

Usage

from datasets import load_dataset

ds = load_dataset("abdelstark/latent-inspector-fingerprints", "fingerprints", split="elephant")
print(ds.to_pandas())

pairs = load_dataset("abdelstark/latent-inspector-fingerprints", "comparisons", split="elephant")
print(pairs.to_pandas())

Or load the raw report directly:

import json, urllib.request
url = "https://huggingface.co/datasets/abdelstark/latent-inspector-fingerprints/resolve/main/compare.json"
data = json.loads(urllib.request.urlopen(url).read())
print(data["overview"]["model_highlights"])

Provenance

• Input image: canonical elephant sample (elephant_sample_image.jpg in the latent-inspector repo)
• Preprocessing: each model's own required transform (224 for EUPE, 256 for the rest), ImageNet normalization
• Backend: ONNX Runtime for all four models, running the three ONNX exports from this account (V-JEPA 2 fpc2, V-JEPA 2 img16, EUPE) plus DINOv2 ViT-L/14 and I-JEPA ViT-H/14
• Tool: latent-inspector compare (source)
• Report timestamp: 2026-04-08

Reproducing

cargo install --git https://github.com/AbdelStark/latent-inspector latent-inspector
latent-inspector compare elephant.jpg \
  --models dinov2-vit-l14,ijepa-vit-h14,vjepa2-vitl-img16-256,eupe-vit-b16 \
  --output ./compare-out \
  --format json

Limitations

• One image. This is a fingerprint, not a benchmark. CKA and k-NN overlap values will drift across images; treat them as a signature, not an absolute ranking.
• EUPE vs ViT-L family has different patch grids (196 vs 256); cross-comparisons truncate to the shared 196 tokens.
• CLS cosine is mostly null because only DINOv2 and EUPE expose CLS tokens and they're dim-incompatible (1024 vs 768).

License

MIT. The underlying model weights keep their upstream licenses (CC-BY-NC-4.0 for DINOv2 / I-JEPA / V-JEPA 2 — research use; FAIR noncommercial research license for EUPE). This dataset only contains derived summary statistics, not model weights.