README.md

metadata

pretty_name: SCIDOCS multivector embeddings (ColBERTv2)
task_categories:
  - text-retrieval
tags:
  - late-interaction
  - colbert
  - multivector
  - beir
  - scidocs

SCIDOCS, ColBERTv2

Token-level (late-interaction) embeddings of the BEIR SCIDOCS corpus and queries, encoded with ColBERTv2, in the TACHIOM multivector format.

Source

• BEIR SCIDOCS, test (its only split) split. Corpus, queries and qrels were read from the official BEIR files via ir_datasets (beir/scidocs); PyLate only did the encoding
• 25,657 documents, 1,000 queries, 29,928 qrels
• Text given to the encoder for each document: title + " " + text (BEIR title and body joined by a space). The text itself is not included, only its vectors
• Row order follows the BEIR corpus and query files; row i of doc_ids.npy / queries_ids.npy identifies row i of doclens.npy / queries.npy

Encoding

• Model: colbert-ir/colbertv2.0 @ c1e84128e85ef755c096a95bdb06b47793b13acf
• Library: PyLate 1.6.0, CPU
• Document length cap: 180 tokens (model default)
• Query length: 32 tokens (model default)
• Query expansion: yes (model default)
• The model defaults come from artifact.metadata in the model repository. We did not override any of them
• Documents: punctuation tokens are dropped (PyLate skiplist), as are padding tokens
• Vectors: 128-d, L2-normalized

Statistics

Token vectors (N)	3,770,957
Avg vectors per document	147.0 (max 180)
Vectors per query	always 32 (padded with [MASK] expansion tokens, which are real embeddings)
Avg vectors per query	32.0

Files

File	dtype	Shape	Content
documents.npy	float16 (<f2)	[3770957, 128]	All document vectors, concatenated document by document
doclens.npy	int32	[25657]	Vectors per document; sum == N
token_ids.npy	uint32	[3770957]	Input token id of each row of documents.npy
doc_ids.npy	string	[25657]	BEIR doc id of each document
queries.npy	float32	[1000, 32, 128]	Query vectors, zero-padded at the end
query_lens.npy	int32	[1000]	True number of vectors per query
queries_ids.npy	string	[1000]	BEIR query id of each query
qrels.test.tsv	TREC	29928 lines	qid \t 0 \t docid \t relevance
groundtruth/gt_top100.tsv	TSV	100000 lines	Exhaustive top-100: query_idx \t doc_idx \t rank \t score (0-based positions)
groundtruth/gt_ids.npy	int32	[1000, 100]	Same, as doc positions
groundtruth/gt_scores.npy	float32	[1000, 100]	Same, MaxSim scores

Zero padding does not change any score: a zero query vector adds 0 to the MaxSim of every document.

Exhaustive-search effectiveness

Exact MaxSim over the full collection (vectorium compute_groundtruth_multivec). These are the reference numbers for approximate search on this data.

nDCG@10	R@100
0.1581	0.3578