This repository contains the development data files used in the SISAP2025 indexing challenge.

Datasets

Common Crawl CCNEWS:
- repo: https://huggingface.co/datasets/sentence-transformers/ccnews/
- Sentence BERT model: sentence-transformers/all-MiniLM-L6-v2
- similarity: Cosine / dot product
- dimension: 384
- corpus size: 614,664
- fields: title and article, train embeddings come from article.
Google Questions & Answers:
- repo: https://huggingface.co/datasets/sentence-transformers/gooaq/
- Sentence BERT model: sentence-transformers/all-MiniLM-L6-v2
- similarity: Cosine / dot product
- dimension: 384
- corpus size: 3,012,496
- fields: questions and answers, train embeddings come from answers.
PUBMED23 from MegRag:
- repo: https://huggingface.co/datasets/MedRAG/pubmed
- Sentence BERT model: sentence-transformers/all-MiniLM-L6-v2
- similarity: Cosine / dot product
- dimension: 384
- corpus size: 23,898,701
- fields: title and content, train embeddings come from content.
Yahoo answers:
- repo: https://huggingface.co/datasets/sentence-transformers/yahoo-answers
- Sentence BERT model: sentence-transformers/all-MiniLM-L6-v2
- similarity: Cosine / dot product
- dimension: 384
- corpus size: 681,164
- fields: question and answer, train embeddings come from answer.

Tasks

Task 1: $k=30$ nearest neighbor search on out-of-distribution queries in PUBMED23. -- sentence bert embeddings sentence-transformers/all-MiniLM-L6-v2 from https://huggingface.co/datasets/MedRAG/pubmed.
Task 2: $k=15$ nearest neighbor graph (a.k.a. self-join problem) on Google QA -- sentence bert embeddings sentence-transformers/all-MiniLM-L6-v2 from https://huggingface.co/datasets/sentence-transformers/gooaq/

ALL of these files correspond to development files. The evaluation process will use other querysets and subsets.

Additionally to PUBMED23 and Google QA we provide benchmark files for CCNEWS and Yahoo QA to help on tuning and experimentation.

h5 datasets

The benchmark files are provided in h5 format and contain datasets and attributes

🗂️  HDF5.File: (read-only) benchmark-dev-pubmed23.h5
├─ 📂 itest
│  ├─ 🏷️  algo
│  ├─ 🏷️  querytime
│  ├─ 🔢 dists
│  ├─ 🔢 knns
│  └─ 🔢 queries
├─ 📂 otest
│  ├─ 🏷️  algo
│  ├─ 🏷️  querytime
│  ├─ 🔢 dists
│  ├─ 🔢 knns
│  └─ 🔢 queries
└─ 🔢 train

Where

train is the main dataset to be indexed (a $384 \times n$ matrix)
otest/queries is the out-of-distribution query set (a $384 \times n$ matrix)
otest/knns (identifiers starting at 1) and otest/dists (distances) are the pre-computed gold standard for the out-of-distribution queries ($384 \times 11000$ matrices)
attributes of otest:
- algo the algorithm used to create the gold-standard.
- querytime the time in seconds used to compute the gold standard (using an Intel(R) Xeon(R) Silver 4216 CPU @ 2.10GHz with 60 threads)
itest/queries is the in-distribution query set, useful for testing and comparing with the out-of-distribution ones (a $384 \times n$ matrix)
itest/knns (identifiers starting at 1) and itest/dists (distances) are the pre-computed gold standard for the in-distribution queries ($384 \times 11000$ matrices)
attributes of itest:
- algo the algorithm used to create the gold-standard.
- querytime the time in seconds used to compute the gold standard (using an Intel(R) Xeon(R) Silver 4216 CPU @ 2.10GHz with 60 threads)

Task2 h5 files

benchmark-dev-*.h5 files can be used for development and both gold-standards follow task1. For task2 we provide an additional gold-standard allknn-benchmark-dev-gooaq.h5 that has the following structure:

🗂️  HDF5.File: (read-only) allknn-benchmark-dev-gooaq.h5
├─ 🏷️  algo
├─ 🏷️  querytime
├─ 🔢 dists 
└─ 🔢 knns

The descriptions of these h5 datasets are similar to the previous ones, however:

knns (identifiers) and dists (distances) corresponding to the outgoing vertices of the $k$ nearest neighbor graph, i.e., the adjacency list of the graph. (a 16 \times n matrix)
the edges contain a self loop that will be ignored by our scoring function.