MERU-ImageNet-1M / README.md
YoKONCy's picture
Upload 2 files
f8781d0 verified
|
Raw History Blame Contribute Delete
10.1 kB
metadata
pretty_name: MERU-ImageNet-1M
language:
  - en
license: cc-by-nc-4.0
size_categories:
  - 1M<n<10M
tags:
  - arxiv:2611.22159
  - hyperbolic
  - approximate-nearest-neighbor
  - vector-search
  - imagenet
  - meru
  - poincare
  - lorentz
  - benchmark

MERU-ImageNet-1M

A Million-Scale Hyperbolic Vector Search Benchmark

MERU-ImageNet-1M contains 1,000,000 database vectors, 10,000 test queries, and 2,000 tuning queries in 512-dimensional hyperbolic space. Astrmira Tech prepared the benchmark by encoding ILSVRC2012 images with the official MERU ViT-small checkpoint and computing exhaustive FP64 nearest-neighbor ground truth.

This is the benchmark called MERU-1M in Pelorus: Warped-Product Graph Indexing for Hyperbolic Vector Search.

The release provides canonical Poincaré coordinates at curvature −1, the model's original Lorentz spatial components, top-10 and top-100 neighbors, distances, image provenance records, and the frozen preparation scripts.

Splits

Split Source images Vectors Dimensions
Database ILSVRC2012 training images 1,000,000 512
Test queries ILSVRC2012 validation images 10,000 512
Tuning queries ILSVRC2012 validation images 2,000 512

The complete training and validation archives were encoded before subset selection. Image IDs, image SHA256 hashes, and canonical vector-byte SHA256 hashes were deduplicated globally. A fixed random seed of 20261114 identifies the preparation run. The selected splits have disjoint image IDs, image hashes, and vector hashes.

Download and load

pip install huggingface_hub numpy
hf download YoKONCy/MERU-ImageNet-1M --repo-type dataset --local-dir MERU-ImageNet-1M
import sys
from pathlib import Path

root = Path("MERU-ImageNet-1M")
sys.path.insert(0, str(root.resolve()))
from load_dataset import load_benchmark

data = load_benchmark(root)
database = data["documents"]
queries = data["queries"]
truth = data["truth"]
print(database.shape, queries.shape, truth.shape)

The arrays are read-only NumPy memory maps. The canonical database occupies 2,048,000,000 bytes. The complete release is approximately 4.32 GB, including the original model-space representation and provenance.

For a selective download of the canonical benchmark, use the following files with snapshot_download.

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="YoKONCy/MERU-ImageNet-1M",
    repo_type="dataset",
    local_dir="MERU-ImageNet-1M",
    allow_patterns=[
        "documents.f32", "queries.f32", "tuning_queries.f32",
        "truth.u64", "tuning_truth.u64", "mapping.json",
        "load_dataset.py", "README.md", "LICENSE*", "SHA256SUMS",
    ],
)

File formats

All binary arrays are headerless, little-endian, and stored in row-major order. Neighbor IDs are zero-based row indices into documents.f32.

File Shape Element type Contents
documents.f32 1,000,000 × 512 float32 Canonical database coordinates
queries.f32 10,000 × 512 float32 Canonical test queries
tuning_queries.f32 2,000 × 512 float32 Canonical tuning queries
truth.u64 10,000 × 10 uint64 Test top-10 neighbor IDs
tuning_truth.u64 2,000 × 10 uint64 Tuning top-10 neighbor IDs
queries_top100.u64 10,000 × 100 uint64 Test top-100 neighbor IDs
queries_top100_distances.f64 10,000 × 100 float64 Test top-100 geodesic distances
tuning_queries_top100.u64 2,000 × 100 uint64 Tuning top-100 neighbor IDs
tuning_queries_top100_distances.f64 2,000 × 100 float64 Tuning top-100 geodesic distances
documents_model_spatial.f32 1,000,000 × 512 float32 Original Lorentz spatial components
queries_model_spatial.f32 10,000 × 512 float32 Original test Lorentz spatial components
tuning_queries_model_spatial.f32 2,000 × 512 float32 Original tuning Lorentz spatial components

The three *_manifest.jsonl.gz files associate vector rows with image IDs, source splits, TAR members, archive offsets, image MD5/SHA256 hashes, and canonical vector SHA256 hashes. Database provenance also includes ImageNet synset IDs.

mapping.json records the model checkpoint hash, source commit, curvature, software versions, source archive hashes, preparation parameters, and binary-file hashes. splits_complete.json records split statistics and database class counts. verification.json and provenance/data_audit.json preserve the ground-truth and integrity checks.

SHA256SUMS covers every published file other than the checksum list itself.

Geometry

Let $y\in\mathbb{R}^{512}$ denote the stored model-space Lorentz spatial component and let $c=0.10000000149011612$ be the original positive curvature magnitude. The canonical Poincaré coordinates are

z=c y,x=z1+∥z∥2+1. z=\sqrt{c}\,y,\qquad x=\frac{z}{\sqrt{1+\lVert z\rVert^2}+1}.

The conversion uses FP64 arithmetic followed by an FP32 cast. Every published canonical vector lies strictly inside the unit ball.

Canonical geodesic distance is

d(x,q)=arcosh⁡ ⁣(1+2∥x−q∥2(1−∥x∥2)(1−∥q∥2)). d(x,q)=\operatorname{arcosh}\!\left( 1+\frac{2\lVert x-q\rVert^2} {(1-\lVert x\rVert^2)(1-\lVert q\rVert^2)} \right).

The corresponding distance at the original model curvature is $d(x,q)/\sqrt{c}$. The shared scale factor preserves nearest-neighbor rankings. Stored top-100 distances use canonical curvature −1 and the published FP32 canonical coordinates.

Ground truth and evaluation

Every test and tuning query was exhaustively compared against all 1,000,000 database vectors using FP64 arithmetic. Neighbors are ordered by increasing distance, with database row ID breaking exact ties. The top-10 files contain the first ten IDs from the corresponding top-100 files.

An independent CPU full scan using direct coordinate differences reproduced all top-100 IDs for 16 fixed test queries. Its maximum distance discrepancy was approximately $2.57\times10^{-15}$. Additional release checks cover file hashes, finite coordinates, ball membership, split disjointness, neighbor ID ranges, ordering, and stored distances.

For ANN evaluation, compute Recall@10 as the mean fraction of the ten ground-truth IDs returned for each test query. Select search parameters using the tuning queries. Report the distance convention, query concurrency, build parameters, latency, and throughput alongside recall.

The benchmark measures geometric nearest-neighbor retrieval under the MERU-ImageNet embedding distribution. Database Euclidean radii range from approximately 0.11673 to 0.13501. splits_complete.json provides the complete recorded quantiles.

Provenance and reproduction

  • Encoder family is MERU ViT-small, with 512 output dimensions.
  • MERU source commit is 6d4430055f58354539f68668175135256f99d47f.
  • Checkpoint SHA256 is 4466c04a790d9b5974928e7e9e528cbff2d82348f596e60443a88934b9a61de6.
  • Encoding uses FP32, with TF32 disabled, batch size 128, and eight image-decoding workers.
  • ILSVRC2012 training archive MD5 is 1d675b47d978889d74fa0da5fadfb00e.
  • ILSVRC2012 validation archive MD5 is 29b22e2961454d5413ddabcf34fc5622.

The original preparation scripts and their recorded hashes are included under reproduction/. The reproduction guide describes their Linux/CUDA environment and source-directory conventions. Source image archives are obtained through ImageNet's access process. MERU project, official checkpoint, ImageNet.

License and attribution

Astrmira Tech releases its benchmark curation under CC BY-NC 4.0. The release retains MERU's CC BY-NC 4.0 notice and ImageNet's non-commercial research and educational use terms. The upstream materials remain subject to their respective rights and access terms. See LICENSE.md, licenses/MERU-LICENSE.txt, and the upstream sources. MERU license, ImageNet terms.

Citation

If you use MERU-ImageNet-1M in your work, please cite the Pelorus paper.

@misc{xiao2026pelorus,
  title         = {{Pelorus}: Warped-Product Graph Indexing for Hyperbolic Vector Search},
  author        = {Xiao, Wenxuan and Cao, Xu},
  year          = {2026},
  eprint        = {2611.22159},
  archivePrefix = {arXiv},
  primaryClass  = {cs.DB},
  doi           = {10.48550/arXiv.2611.22159},
  url           = {https://arxiv.org/abs/2611.22159}
}

For the dataset release, the MERU encoder, and the ImageNet source images, the corresponding references are listed below.

@misc{astrmira2026meruimagenet1m,
  title        = {MERU-ImageNet-1M — A Million-Scale Hyperbolic Vector Search Benchmark},
  author       = {Xiao, Wenxuan and Cao, Xu},
  year         = {2026},
  version      = {1.0},
  publisher    = {Astrmira Tech},
  howpublished = {Hugging Face},
  url          = {https://huggingface.co/datasets/YoKONCy/MERU-ImageNet-1M}
}

@inproceedings{desai2023meru,
  title     = {Hyperbolic Image-Text Representations},
  author    = {Desai, Karan and Nickel, Maximilian and Rajpurohit, Tanmay and Johnson, Justin and Vedantam, Ramakrishna},
  booktitle = {Proceedings of the International Conference on Machine Learning},
  year      = {2023}
}

@article{russakovsky2015imagenet,
  title   = {ImageNet Large Scale Visual Recognition Challenge},
  author  = {Russakovsky, Olga and Deng, Jia and Su, Hao and Krause, Jonathan and Satheesh, Sanjeev and Ma, Sean and Huang, Zhiheng and Karpathy, Andrej and Khosla, Aditya and Bernstein, Michael and Berg, Alexander C. and Fei-Fei, Li},
  journal = {International Journal of Computer Vision},
  year    = {2015},
  volume  = {115},
  pages   = {211--252}
}