PIIR/Iceberg-dataset

Iceberg: Task-Centric Benchmarks for Vector Similarity Search

The Iceberg benchmark was presented in the paper Reveal Hidden Pitfalls and Navigate Next Generation of Vector Similarity Search from Task-Centric Views.

Code Repository: https://github.com/ZJU-DAILY/Iceberg

Introduction

Iceberg is a comprehensive benchmark suite for end-to-end evaluation of VSS (Vector Similarity Search) methods in realistic application settings. From a task-centric view, Iceberg uncovers the Information Loss Funnel, which identifies three principal sources of end-to-end performance degradation: (1) Embedding Loss during feature extraction; (2) Metric Misuse, where distances poorly reflect task relevance; (3) Data Distribution Sensitivity, highlighting index robustness across skews and modalities.

Iceberg spans 7 diverse datasets across key domains including image classification, face recognition, text retrieval, and recommendation systems. Each dataset contains 1M to 100M vectors enriched with task-specific labels and metrics, enabling evaluation of retrieval algorithms within full application pipelines—not just in isolated recall-speed scenarios. Iceberg benchmarks 13 state-of-the-art VSS algorithms and re-ranks them using task-centric performance metrics, uncovering substantial deviations from conventional recall/speed-based rankings. Morever, Iceberg propose an interpretable decision tree to guide practitioners in selecting and tuning VSS methods for specific workloads.

Datasets

Overview

Dataset	Base Size	Dim	Query Size	Domain	Origin data source
ImageNet-DINOv2	1,281,167	768	50,000	Image Classification	https://image-net.org/index.php
ImageNet-EVA02	1,281,167	1024	50,000	Image Classification	https://image-net.org/index.php
ImageNet-ConvNeXt	1,281,167	1536	50,000	Image Classification	https://image-net.org/index.php
Glink360K-IR101	17,091,649	512	20,000	Face Recognition	https://github.com/deepinsight/insightface/tree/master/recognition/partial_fc#glint360k
Glink360K-ViT	17,091,649	512	20,000	Face Recognition	https://github.com/deepinsight/insightface/tree/master/recognition/partial_fc#glint360k
BookCorpus	9,250,529	1024	10,000	Text Retrieval	https://huggingface.co/datasets/bookcorpus/bookcorpus
Commerce	99,085,171	48	64,111	Recommendation

Detailed Description

D1: ImageNet

ImageNet is a large-scale dataset containing millions of high-resolution images spanning thousands of object categories. Each image is annotated with ground-truth labels, either manually or semi-automatically. The dataset has been widely used in the computer vision community for model training and benchmarking, particularly for image classification tasks.

Embedding Models:

• DINOv2: https://huggingface.co/facebook/dinov2-base
• EVA02: https://huggingface.co/timm/eva02_large_patch14_448.mim_m38m_ft_in22k_in1k
• ConvNeXt: https://huggingface.co/timm/convnext_large_mlp.clip_laion2b_soup_ft_in12k_in1k_384

End Tasks:

• Label Recall@K: It measures how many correct task-specific labels appear in the top-K retrieved results.

D2: Glink360K

Glint360K is a large-scale face dataset created by merging and cleaning multiple public face datasets to significantly expand both the number of identities and facial images.

Embedding Models:

• Resnet-IR101: https://huggingface.co/minchul/cvlface_arcface_ir101_webface4m
• ViT: https://huggingface.co/gaunernst/vit_tiny_patch8_112.arcface_ms1mv3

End Tasks:

• Label Recall@K: It measures how many correct task-specific labels appear in the top-K retrieved results.

D3: BookCorpus

BookCorpus consists of text extracted from approximately 19,000 books spanning various domains and has been curated into a high-quality corpus. The text was segmented at the paragraph level, with each paragraph concatenated into chunks containing eight sentences. This preprocessing resulted in a base dataset of 9,250,529 paragraphs. From this corpus, 10,000 paragraphs were randomly sampled to construct the query set. The unique ID of each paragraph was used as the label for its corresponding embedding vector.

Embedding Models:

• Stella: https://huggingface.co/NovaSearch/stella\_en\_1.5B\_v5

End Tasks:

• Hit@K: It measures whether the most semantic relevant paragraph is included in the top-K retrieved results.

D4: Commerce

Commerce dataset, derived from anonymized traffic logs of a major e-commerce platform, serves as a representative benchmark for large-scale E-commerce systems. Collected over several months, the dataset comprises 99,085,171 records of frequently purchased grocery items. In addition, a query set of 64,111 entries was constructed to represent user profiles and associated search keywords. Each query is linked to a sequence of high-popularity items, enabling evaluation on downstream recommendation tasks. Item IDs are used as labels throughout the dataset.

Embedding Models:

• ResFlow: https://github.com/FuCongResearchSquad/ResFlow

End Tasks:

• Matching Score@K: It measures whether the vectors retrieved by a query are both relevant and popular, as well as the cumulative popularity of those items.

Supported Algorithms

	Metric	Type	Original Code Link
Fargo	Inner Product	Parition-based	https://github.com/Jacyhust/FARGO_VLDB23
ScaNN	Inner Product	Parition-based	https://github.com/google-research/google-research/tree/master/scann
ip-NSW	Inner Product	Graph-based	https://github.com/stanis-morozov/ip-nsw
ip-NSW+	Inner Product	Graph-based	https://github.com/jerry-liujie/ip-nsw/tree/GraphMIPS
Mobius	Inner Product	Graph-based	Our own implementation
NAPG	Inner Product	Graph-based	Our own implementation
MAG	Inner Product	Graph-based	https://github.com/ZJU-DAILY/MAG
RaBitQ	Euclidean Distance	Parition-based	https://github.com/VectorDB-NTU/RaBitQ-Library
IVFPQ	Euclidean Distance	Parition-based	https://github.com/facebookresearch/faiss
DB-LSH	Euclidean Distance	Parition-based	https://github.com/Jacyhust/DB-LSH
HNSW	Euclidean Distance	Graph-based	https://github.com/nmslib/hnswlib
NSG	Euclidean Distance	Graph-based	https://github.com/ZJULearning/nsg
Vamana	Euclidean Distance	Graph-based	https://github.com/microsoft/DiskANN

Quick Start

Clone the repository

git clone project

Environment Requirements

Python 3.10+; docker; pyyaml

Run pip install -r requirements.txt.

Run the benchmark

Example: We use HNSW for the ImageNet dataset as an example to run the benchmark.

• Configure the dataset (config/dataset.yaml):

  imagenet1k_avg:
    dataset_type: imagenet
    data_pre: imagenet-1k
    train_name: convnext-avg-pool-train.bin
    test_name: convnext-avg-pool-validation.bin
    train_path: /workspace/data/imagenet-1k/convnext-avg-pool-train.bin
    test_path: /workspace/data/imagenet-1k/convnext-avg-pool-validation.bin
    prefix: convnext-avg-pool
    data_dim: 1536
    k: 100
    data_num: 1281167
    query_num: 50000
  

• Configure the algorithm (config/algorithm.yaml)

  hnsw:
    efc: 256
    M: 32
    efs: [100, 200, 300, 400, 500, 600, 800, 1000, 1500]
    type: nn
  

  Configuration parameters:

  • efc: build parameter for HNSW
  • M: build parameter for HNSW
  • efs: search parameter for HNSW
  • type: distance metric type

• run the algorithm & evaluation

  1. Configure the dataset and algorithm parameters in config/dataset.yaml and config/algorithm.yaml
  2. Run the algorithm using: python3 run.py hnsw imagenet1k_dinov2 --mode build/search
  3. For more configuration options, refer to: python run.py --help

Pipeline

Results

Iceberg LeaderBoard 1.0

Task-centric performance versus two similarity metrics

Query Performance on Synthetic Recall@100

Citation

@article{chen2025iceberg,
  title={Reveal Hidden Pitfalls and Navigate Next Generation of Vector Similarity Search from Task-Centric Views},
  author={Chen, Tingyang and Fu, Cong and Wu, Jiahua and Wu, Haotian and Fan, Hua and Ke, Xiangyu and Gao, Yunjun and Ni, Yabo and Zeng, Anxiang},
  journal={arXiv preprint arXiv:2512.12980},
  year={2025},
  url={https://arxiv.org/abs/2512.12980},
}