tuskanny/sparse-msmarco-efficient-splade

MS MARCO passages, Efficient-SPLADE

Learned sparse embeddings of the MS MARCO v1 passage corpus and its dev (small) queries, encoded with Efficient-SPLADE (V-large), ready for approximate nearest neighbor search by inner product. The query encoder is trained to produce queries with few nonzero entries, so queries are much shorter than documents.

Source

• MS MARCO v1 passage ranking: 8,841,823 passages, 6,980 queries (dev, small)
• Models: naver/efficient-splade-V-large-doc for documents, naver/efficient-splade-V-large-query for queries
• Only the vectors are included, not the text

Statistics

Documents	8,841,823
Queries	6,980
Avg nonzeros per document	182.8
Avg nonzeros per query	5.9
Terms in token_to_id_mapping.json	29,452

Files

File	Size	Content
documents.bin	13.0 GB	Document vectors
queries.bin	360 kB	Query vectors
groundtruth.tsv	1.7 MB	Exact top-10 per query
doc_ids.npy	71 MB	MS MARCO passage id of each document (int64)
queries_ids.npy	56 kB	MS MARCO query id of each query (int64)
token_to_id_mapping.json	510 kB	Term to component id

Format

Seismic binary format, little-endian, used by Seismic, kANNolo and Vectorium: u32 n_vectors, then for each vector u32 nnz, nnz x u32 component ids (ascending), nnz x f32 values.

groundtruth.tsv has four tab-separated columns: query position, document position, rank (1-based), score. Positions are 0-based rows of queries.bin and documents.bin. It holds the exact top-10 of every query by inner product, computed with an exhaustive scan. Row i of doc_ids.npy / queries_ids.npy is the original id of row i of documents.bin / queries.bin.

Component ids are collection-specific term ids; token_to_id_mapping.json maps terms to component ids.