README.md

metadata

pretty_name: NFCorpus multivector embeddings (LateOn)
task_categories:
  - text-retrieval
tags:
  - late-interaction
  - colbert
  - multivector
  - beir
  - nfcorpus

NFCorpus, LateOn

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

Source

• BEIR NFCorpus, test split. Corpus, queries and qrels were read from the official BEIR files via ir_datasets (beir/nfcorpus/test); PyLate only did the encoding
• 3,633 documents, 323 queries, 12,334 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: lightonai/LateOn @ 62911e105059585d244384c7d17826e35f669c17
• Library: PyLate 1.6.0, CPU
• Document length cap: 300 tokens (model default)
• Query length: 32 tokens (model default)
• Query expansion: no (model default)
• The model defaults come from config_sentence_transformers.json 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)	862,599
Avg vectors per document	237.4 (max 286)
Vectors per query	variable, 4 to 25 (no query expansion), zero-padded to 25
Avg vectors per query	8.6

Files

File	dtype	Shape	Content
documents.npy	float16 (<f2)	[862599, 128]	All document vectors, concatenated document by document
doclens.npy	int32	[3633]	Vectors per document; sum == N
token_ids.npy	uint32	[862599]	Input token id of each row of documents.npy
doc_ids.npy	string	[3633]	BEIR doc id of each document
queries.npy	float32	[323, 25, 128]	Query vectors, zero-padded at the end
query_lens.npy	int32	[323]	True number of vectors per query
queries_ids.npy	string	[323]	BEIR query id of each query
qrels.test.tsv	TREC	12334 lines	qid \t 0 \t docid \t relevance
groundtruth/gt_top100.tsv	TSV	32300 lines	Exhaustive top-100: query_idx \t doc_idx \t rank \t score (0-based positions)
groundtruth/gt_ids.npy	int32	[323, 100]	Same, as doc positions
groundtruth/gt_scores.npy	float32	[323, 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
0.3809	0.3366