Repository navigation
Hope you can recommend me more datasets #2838
Replies: 2 comments
|
A few standard sources for nearest-neighbor benchmarking, by scale: Million-scale, ready-made with ground truth:
Billion-scale:
Text-embedding (if you want semantic rather than synthetic vectors):
For algorithm-effectiveness work, the ANN-Benchmarks suite is usually the fastest starting point because the ground truth and the evaluation harness are already provided. |
To add to the above section, if you are testing an algorithm for modern production systems, it is highly recommended to benchmark against the four major workloads that dominate current Approximate Nearest Neighbor (ANN) research: 1. Filtered Search (FANNS)Real-world vector search almost always includes metadata filters (e.g., "find the nearest neighbor where category = clothing and price < $50"). Many fast algorithms fail when hard filters are applied because the vector space becomes fragmented.
2. Streaming and Dynamic VectorsModern vector databases must handle continuous insertions, deletions, and updates without degrading search recall or rebuilding the entire index.
3. Out-of-Distribution (OOD) & Cross-ModalIn real systems, the distribution of query vectors often doesn't match the distribution of the base database (e.g., querying an image database using text embeddings). Algorithms that overfit to the base distribution often fail here.
4. Sparse and Hybrid SearchModern retrieval-augmented generation (RAG) systems combine dense distance metrics with high-dimensional sparse vectors (like BM25 or SPLADE) calculated via Maximum Inner Product Search (MIPS).
|
Uh oh!
There was an error while loading. Please reload this page.
Hi, I am a master's student studying nearest neighbor search algorithm. At present, we need more datasets to test the effectiveness of the algorithm. I hope you can recommend me more dataset links and other information. You can contact me at “tragrace0125@gmail.com”. Thank you very much!
All reactions