Datasets:

Comet
/

wikipedia-2017-bm25

license: cc-by-sa-3.0
task_categories:
  - text-retrieval
language:
  - en
tags:
  - bm25
  - wikipedia
  - information-retrieval
  - research
size_categories:
  - 1M<n<10M

Wikipedia 2017 BM25 Search Index

This dataset provides a production-ready BM25 search index over 5.2 million Wikipedia article abstracts from the 2017 snapshot. Built using the bm25s library with English stemming and optimized Parquet compression, it enables fast, offline information retrieval for research and production AI systems. The corpus is identical to the one used in influential AI research papers including DSPy and GEPA, ensuring reproducible benchmarking and fair comparison across studies.

The index uses BM25 (Best Matching 25), a probabilistic ranking function widely recognized as the gold standard for lexical search. With carefully tuned parameters (k1=0.9 for term frequency saturation, b=0.4 for document length normalization), it provides state-of-the-art retrieval performance for factual queries. Each search completes in under 100ms on consumer hardware, making it suitable for real-time applications, RAG (Retrieval-Augmented Generation) pipelines, and agent tool implementations.

We created this index for the Opik Optimizer project to enable reproducible prompt optimization experiments and agent benchmarking. By using the same Wikipedia 2017 corpus as established research, we ensure that optimization results are directly comparable to published baselines. The Parquet-compressed format reduces download size by 67% while maintaining full search fidelity, making it practical for cloud deployments and CI/CD pipelines where storage costs and download times matter.

Size: 1.61 GB | Format: Parquet (67% compressed) | Documents: 5.2M

Format: Optimized (Parquet)

This is the optimized version with 40-50% size reduction:

Corpus Format: Chunked Parquet with ZSTD compression (level 9)
Index Format: NumPy compressed arrays (.npz)
Size: ~2.5-3.5 GB (vs ~4.87 GB standard)

Benefits:

Smaller downloads: 40-50% reduction vs standard format
Streaming access: Load only needed document chunks
Faster HF downloads: Parallel chunk downloads
Better compression: ZSTD level 9 on columnar format

When to use:

Running on Modal or cloud workers (storage costs matter)
Bandwidth-constrained environments
High-volume deployments

Quick Start

Install package

pip install opik_optimizer[bm25]

Run the search

from opik_optimizer.utils.tools.wikipedia import search_wikipedia

results = search_wikipedia(
    "quantum entanglement",
    search_type="bm25",
    n=5,
    bm25_hf_repo="Comet/wikipedia-2017-bm25"
)

That's it! First run downloads the index (~1.6 GB), subsequent searches are instant.

Why Use This?

Reproducible Research - Same corpus used in various prompt and agent optimization papers
Fast & Offline - No API rate limits, <100ms query time
Production Ready - Powers RAG systems, Q&A benchmarks, agents
Memory Efficient - Optimized Parquet format with chunked loading

Use Cases

Research & Benchmarking

HotpotQA multi-hop question answering
Information retrieval experiments
RAG pipeline evaluation
Agent tool development

Production Applications

Offline knowledge base for AI agents
Research paper search
Educational tools
Content recommendation

Index Specifications

Attribute	Value
Documents	5,233,330 Wikipedia abstracts
Source	Wikipedia 2017 dump (DSPy cache)
Tokenization	English stemming + stopword removal
Algorithm	BM25 (k1=0.9, b=0.4)
Library	`bm25s`
Memory	~6-8 GB RAM during search
Query Speed	<100ms per search

License & Attribution

Dataset License: CC-BY-SA 3.0 (Wikipedia content license) Source: Wikipedia 2017 abstracts from DSPy cache

Citation: If you use this in research, please cite:

@misc{comet_ml_wikibm25,
    author       = { Comet ML and Vincent Koc },
    title        = { wikipedia-2017-bm25 (Revision 07db6d6) },
    year         = 2025,
    url          = { https://huggingface.co/datasets/Comet/wikipedia-2017-bm25 },
    doi          = { 10.57967/hf/7073 },
    publisher    = { Hugging Face }
}