PaDaS-Lab/ecir2027-curse-dense-retrieval-run-files

Full-corpus run files: Revisiting Dense Information Retrieval for Large Index Sizes

This dataset contains the retrieval run files (top 1,000 documents per query) of the ECIR 2027 reproducibility paper "Revisiting Dense Information Retrieval for Large Index Sizes". They were produced by BM25, SPLADE and three uncompressed dense retrievers on the full (not subsampled) corpora used in the paper. The code that produced and consumes them is the code repository linked in the paper.

Layout

The folder structure mirrors the RUNS_FOLDER of the code repository exactly:

<dataset>/full/<num_docs>/<system>.run.gz

<dataset> is the ir-datasets id with / replaced by _, <num_docs> is the number of documents in the corpus, and <system> is one of bm25, bm25-elasticsearch, splade, dense-sbert, dense-gemma, dense-jinav5. SHA256SUMS lists a checksum for every run file (sha256sum -c SHA256SUMS).

Run format

Each file is a gzipped TREC run: one line per retrieved document, six whitespace-separated columns

qid Q0 docno rank score tag

with at most 1,000 documents per query (rank 1..1000). Sparse (BM25) runs can contain fewer than 1,000 lines for a query if fewer documents match.

Included runs

Dataset (directory)	Corpus size	Queries	Systems
beir_msmarco_dev (MS MARCO v1 dev)	8,841,823	6,980	all six
core_passage (CoRE Passage)	100,000,000	65	all six
core_document (CoRE Document)	10,000,000	55	all six
random-noise_beir_msmarco_dev_random-noise-reimers-no-query-words (Random characters, seed 42)	100,000,000	6,980	all six
random-noise_beir_msmarco_dev_random-noise-reimers-no-query-words-43 (Random characters, seed 43)	100,000,000	6,980	all six
random-noise_beir_msmarco_dev_random-noise-reimers-no-query-words-44 (Random characters, seed 44)	100,000,000	6,980	all six
random-noise_beir_msmarco_dev_random-noise-faker (Random words, seed 42)	100,000,000	6,980	all except dense-gemma
random-noise_beir_msmarco_dev_random-noise-faker-43 (Random words, seed 43)	100,000,000	6,980	all six
random-noise_beir_msmarco_dev_random-noise-faker-44 (Random words, seed 44)	100,000,000	6,980	all six
random-noise_beir_msmarco_dev_random-noise-reimers (verbatim noise)	100,000,000	6,980	all six

Notes:

• dense-gemma was not run on random-noise_beir_msmarco_dev_random-noise-faker (Random words, seed 42).
• bm25-elasticsearch contains 6,979 of the 6,980 queries on the Random characters and Random words corpora (one query has no entry); all other files contain every query.

Not included

• Runs on the subsampled corpora (pooled, random, predefined subsets).
• Runs of the truncated or binarized dense configurations (reduced dimensionality, binary quantization). Their metrics are in results/results.csv of the code repository.
• Runs on auxiliary datasets and systems not used in the paper's tables.

Download

Download directly into the RUNS_FOLDER of the code repository:

hf download PaDaS-Lab/ecir2027-curse-dense-retrieval-run-files --repo-type dataset --local-dir runs/

The download is about 3.2 GB.

Use in the code

Unpacked into RUNS_FOLDER, these files let

• repro pool rebuild the MS MARCO v1 pool used for pooled corpus subsampling, and
• repro significance recompute the significance tests of Table 4.

Systems

• bm25: BM25 with Terrier in Anserini mode (k1=0.9, b=0.4).
• bm25-elasticsearch: BM25 with Elasticsearch (Lucene 8.11, default analyzer, k1=1.2, b=0.75).
• splade: SPLADE v3 (naver/splade-v3).
• dense-sbert: sentence-transformers/msmarco-distilroberta-base-v2.
• dense-gemma: google/embeddinggemma-300m.
• dense-jinav5: jinaai/jina-embeddings-v5-text-nano.

All dense models use uncompressed 768-dimensional embeddings with exact (brute-force) cosine search.

Citation

[TODO: citation, to be added after the double-blind review]