ms_marco_colbertv2

pretty_name: MS MARCO v1 Passage multivector embeddings (ColBERTv2)
task_categories:
  - text-retrieval
tags:
  - late-interaction
  - colbert
  - multivector
  - msmarco

MS MARCO v1 Passage, ColBERTv2

Token-level (late-interaction) ColBERTv2 embeddings of the MS MARCO v1 passage collection and the dev/small queries.

Source

Collection: MS MARCO v1 passage (ir_datasets msmarco-passage), 8,841,823 passages
Queries: dev/small, 6,980 queries and 7,437 qrels (msmarco-passage/dev/small)
Document order: passage id order (row i is pid i)

Encoding

Model: ColBERTv2 (colbert-ir/colbertv2.0, BERT-base-uncased tokenizer)
Documents start with [CLS] and the ColBERT [D] marker (token id 2)
Queries: always 32 vectors ([MASK] expansion, no zero padding)
Vectors: 128-d, L2-normalized
Not recorded: exact checkpoint revision, encoding library, document length cap (the longest document has 300 vectors), and whether punctuation was dropped

Statistics


Token vectors (N)	597,909,919
Avg vectors per document	67.6 (min 4, max 300)
Vectors per query	32

Files

File	dtype	Shape	Content
`documents.npy`	uint16 (`<u2`)	`[597909919, 128]`	Raw float16 bit patterns stored as uint16. Read with `.view(np.float16)`
`doclens.npy`	int32	`[8841823]`	Vectors per document; `sum == N`
`token_ids_per_token.npy`	int64	`[597909919]`	Input token id of each row of `documents.npy`
`doc_ids.npy`	int64	`[8841823]`	MS MARCO pid (equals the row index)
`queries.npy`	float32	`[6980, 32, 128]`	Query vectors
`queries_ids.npy`	int64	`[6980]`	MS MARCO qid of each query

Document ids are MS MARCO passage ids and equal the row index: row i of doclens.npy is passage i.

documents.npy stores float16 values as their raw 16-bit patterns, with dtype uint16. In numpy, read it with np.load("documents.npy", mmap_mode="r").view(np.float16).

Relevance judgments

The official MS MARCO passage dev/small qrels (qrels.dev.small.tsv): 7,437 relevant query-passage pairs over the 6,980 queries, all with relevance 1 (ir_datasets msmarco-passage/dev/small). Standard metric: MRR@10 (RR@10 in ir_measures).