KShivendu/nanobeir-colbert-token-vectors

NanoBEIR ColBERT token vectors

Per-token ColBERT vectors for all 13 NanoBEIR datasets, from two late-interaction models:

folder	model	revision	encoder dtype
lateon/	lightonai/LateOn	62911e105059585d244384c7d17826e35f669c17	fp32
iso/	topk-io/Iso-ModernColBERT	b95d9608ad424fdd7bfd045001576f59cbb89a98	bf16
pplx-late-0.6b/	perplexity-ai/pplx-embed-v2-late-0.6b	dd4e95b836a73f6f0c32e46ea127c0b86b02169e	bf16

Each model covers 56,723 documents (7,585,402 kept document tokens) and 645 queries, 128 dimensions per token. With these files you can run exact MaxSim, build sparse first stages (token-type centroids, SMVE, MUVERA) or train compressors without a GPU.

They were made for a study of cheap first stages for ColBERT (token-type MaxSim against TopK's SMVE, BM25 and cross-encoders).

How they were encoded

• PyLate 1.6.0, sentence-transformers 5.3.0, transformers 5.3.0.
• Prefixes [D] / [Q] , document_length 300, query_length 32, no query expansion.
• Document tokens in the model's punctuation skiplist are removed, as are padding tokens. Query tokens are all kept.
• Document vectors are stored as fp16, query vectors as fp32. Both are the model's normalised outputs, as PyLate returns them.
• Documents were encoded in chunks of 2048 at batch size 128, one dataset at a time.
• Script: export_tokmaxsim.py in the source repo.

Queries longer than 32 tokens are truncated by the model. This matters for ArguAna, whose queries are whole arguments (median 188 words).

pplx-embed-v2-late-0.6b (pplx-late-0.6b/)

• perplexity-ai/pplx-embed-v2-late-0.6b, revision dd4e95b836a73f6f0c32e46ea127c0b86b02169e, MIT license, run in bf16.
• sentence-transformers 6.1.0 MultiVectorEncoder (not PyLate), transformers 5.17.0.
• Prompts [Q] / [D] as the first token, query cap 1,024 tokens, document cap 4,096, no query expansion. The model's mask drops 32 single-character punctuation tokens on documents only.
• Vectors are 128-dim and normalised. Same raw layout and Parquet schema as the other models; token ids use the model's own (Qwen3.5) tokenizer.

Load with datasets

Each model has a corpus and a queries config, with one split per NanoBEIR dataset. A row is one document or query: its text, the kept tokens and one 128-dim vector per kept token.

from datasets import load_dataset
docs = load_dataset("KShivendu/nanobeir-colbert-token-vectors", "iso-corpus", split="NanoNQ")
qs = load_dataset("KShivendu/nanobeir-colbert-token-vectors", "iso-queries", split="NanoNQ")
qrels = load_dataset("KShivendu/nanobeir-colbert-token-vectors", "qrels", split="NanoNQ")
docs[0]["text"], docs[0]["tokens"][:8], len(docs[0]["vectors"])   # vectors: n_tokens x 128

config	columns
iso-corpus, lateon-corpus, pplx-corpus	dataset, doc_id, text, token_ids, tokens, vectors (fp16)
iso-queries, lateon-queries, pplx-queries	dataset, query_id, text, token_ids, tokens, vectors (fp32)
qrels	dataset, query_id, doc_id, score

The Parquet files hold the same vectors as the raw files described below, which are faster to memory-map with NumPy.

Layout

One folder per model and dataset, <model>/<Dataset>/:

file	shape / format	content
meta.json	json	n_doc, n_tok, n_q, n_qtok, dim, model, revision, dtype
dvec.f16	[n_tok, 128] little-endian fp16	document token vectors
doffs.u32	[n_doc + 1] u32	document d owns rows doffs[d] .. doffs[d+1]
dtok.u32	[n_tok] u32	tokenizer id of each kept document token
qvec.f32	[n_qtok, 128] fp32	query token vectors
qoffs.u32 / qtok.u32	u32	query offsets and token ids
ids.txt, qids.txt	one id per line	document and query ids (BEIR / NanoBEIR ids)
qrels.tsv	qid \t doc_id \t grade	relevance labels
ref.json	json	reference top 100 per query for token-type MaxSim, plus the exact-MaxSim rerank
smve_{raw,cen}_{doffs,dims,vals}	CSR per doc	SMVE 65,536-wide / top-32 sparse doc vectors, raw and corpus-mean-centred
exp_doc/exp_id/exp_w	u32/u32/u8	doc-side token-type expansion (cosine >= 0.7, weight x255)
wcen.f32, word_doc.u32, word_id.u32, wkeys.json, wocc_*.u32		whole-word centroids and word occurrences
ref_extra.json	json	reference top 100 for SMVE, expansion and whole-word variants

At the root, smve_B65536.f32 [128, 65536] is the random projection used for SMVE (randn(128, 65536) with seed 65536, unit-norm columns).

Reading one dataset in NumPy:

import json, numpy as np
d = "iso/NanoNQ"
m = json.load(open(f"{d}/meta.json"))
D = np.fromfile(f"{d}/dvec.f16", "<f2").reshape(m["n_tok"], 128)
offs = np.fromfile(f"{d}/doffs.u32", "<u4")
Q = np.fromfile(f"{d}/qvec.f32", "<f4").reshape(m["n_qtok"], 128)
qoffs = np.fromfile(f"{d}/qoffs.u32", "<u4")
q = Q[qoffs[0]:qoffs[1]]                         # first query
maxsim = np.maximum.reduceat(D @ q.T, offs[:-1], axis=0).sum(1)   # score of every doc

To merge all 13 datasets into one corpus, concatenate them and prefix ids with the dataset name (NanoNQ|doc806). The merged corpus holds identical texts under different ids in several datasets (for example NanoClimateFEVER|California and NanoFEVER|California). Only the original id is labelled relevant.

Licenses

These vectors are derived from the NanoBEIR subsets of BEIR datasets. Each source keeps its own license and terms, which vary by dataset and include non-commercial licenses (for example MS MARCO). Check the license of every source dataset before use. The models carry their own licenses on their model pages.

Citations

If you use these vectors, cite the models, NanoBEIR and BEIR, and the source datasets.

@inproceedings{khattab2020colbert,
  title     = {ColBERT: Efficient and Effective Passage Search via Contextualized Late Interaction over BERT},
  author    = {Khattab, Omar and Zaharia, Matei},
  booktitle = {Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval},
  year      = {2020}
}

@misc{lateon,
  title        = {LateOn},
  author       = {{LightOn}},
  howpublished = {\url{https://huggingface.co/lightonai/LateOn}},
  note         = {Revision 62911e105059585d244384c7d17826e35f669c17}
}

@misc{pplxembedv2late,
  title        = {pplx-embed-v2-late-0.6b},
  author       = {{Perplexity AI}},
  howpublished = {\url{https://huggingface.co/perplexity-ai/pplx-embed-v2-late-0.6b}},
  note         = {Revision dd4e95b836a73f6f0c32e46ea127c0b86b02169e. See also https://www.perplexity.ai/hub/blog/multimodal-embeddings-beyond-a-single-vector}
}

@misc{isomoderncolbert,
  title        = {Iso-ModernColBERT},
  author       = {{TopK}},
  howpublished = {\url{https://huggingface.co/topk-io/Iso-ModernColBERT}},
  note         = {Revision b95d9608ad424fdd7bfd045001576f59cbb89a98}
}

@misc{warner2024modernbert,
  title         = {Smarter, Better, Faster, Longer: A Modern Bidirectional Encoder for Fast, Memory Efficient, and Long Context Finetuning and Inference},
  author        = {Warner, Benjamin and Chaffin, Antoine and Clavi{\'e}, Benjamin and Weller, Orion and Hallstr{\"o}m, Oskar and Taghadouini, Said and Gallagher, Alexis and Biswas, Raja and Ladhak, Faisal and Aarsen, Tom and Cooper, Nathan and Adams, Griffin and Howard, Jeremy and Poli, Iacopo},
  year          = {2024},
  eprint        = {2412.13663},
  archivePrefix = {arXiv}
}

@misc{pylate,
  title        = {PyLate},
  author       = {Chaffin, Antoine and Sourty, Rapha{\"e}l},
  howpublished = {\url{https://github.com/lightonai/pylate}},
  note         = {Version 1.6.0}
}

@misc{nanobeir,
  title        = {NanoBEIR},
  author       = {{Zeta Alpha}},
  howpublished = {\url{https://huggingface.co/zeta-alpha-ai}},
  year         = {2024}
}

@inproceedings{thakur2021beir,
  title     = {{BEIR}: A Heterogeneous Benchmark for Zero-shot Evaluation of Information Retrieval Models},
  author    = {Thakur, Nandan and Reimers, Nils and R{\"u}ckl{\'e}, Andreas and Srivastava, Abhishek and Gurevych, Iryna},
  booktitle = {Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2)},
  year      = {2021}
}

@misc{bajaj2016msmarco,
  title         = {{MS MARCO}: A Human Generated MAchine Reading COmprehension Dataset},
  author        = {Bajaj, Payal and Campos, Daniel and Craswell, Nick and Deng, Li and Gao, Jianfeng and Liu, Xiaodong and Majumder, Rangan and McNamara, Andrew and Mitra, Bhaskar and Nguyen, Tri and Rosenberg, Mir and Song, Xia and Stoica, Alina and Tiwary, Saurabh and Wang, Tong},
  year          = {2016},
  eprint        = {1611.09268},
  archivePrefix = {arXiv}
}

@article{kwiatkowski2019nq,
  title   = {Natural Questions: A Benchmark for Question Answering Research},
  author  = {Kwiatkowski, Tom and Palomaki, Jennimaria and Redfield, Olivia and Collins, Michael and Parikh, Ankur and Alberti, Chris and Epstein, Danielle and Polosukhin, Illia and Devlin, Jacob and Lee, Kenton and Toutanova, Kristina and Jones, Llion and Kelcey, Matthew and Chang, Ming-Wei and Dai, Andrew M. and Uszkoreit, Jakob and Le, Quoc and Petrov, Slav},
  journal = {Transactions of the Association for Computational Linguistics},
  volume  = {7},
  pages   = {452--466},
  year    = {2019}
}

@inproceedings{yang2018hotpotqa,
  title     = {{HotpotQA}: A Dataset for Diverse, Explainable Multi-hop Question Answering},
  author    = {Yang, Zhilin and Qi, Peng and Zhang, Saizheng and Bengio, Yoshua and Cohen, William and Salakhutdinov, Ruslan and Manning, Christopher D.},
  booktitle = {Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing},
  year      = {2018}
}

@inproceedings{maia2018fiqa,
  title     = {{WWW'18} Open Challenge: Financial Opinion Mining and Question Answering},
  author    = {Maia, Macedo and Handschuh, Siegfried and Freitas, Andr{\'e} and Davis, Brian and McDermott, Ross and Zarrouk, Manel and Balahur, Alexandra},
  booktitle = {Companion Proceedings of the The Web Conference 2018},
  year      = {2018}
}

@inproceedings{wachsmuth2018arguana,
  title     = {Retrieval of the Best Counterargument without Prior Topic Knowledge},
  author    = {Wachsmuth, Henning and Syed, Shahbaz and Stein, Benno},
  booktitle = {Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics},
  year      = {2018}
}

@inproceedings{bondarenko2020touche,
  title     = {Overview of Touch{\'e} 2020: Argument Retrieval},
  author    = {Bondarenko, Alexander and Fr{\"o}be, Maik and Beloucif, Meriem and Gienapp, Lukas and Ajjour, Yamen and Panchenko, Alexander and Biemann, Chris and Stein, Benno and Wachsmuth, Henning and Potthast, Martin and Hagen, Matthias},
  booktitle = {Working Notes Papers of the CLEF 2020 Evaluation Labs},
  year      = {2020}
}

@misc{diggelmann2020climatefever,
  title         = {{CLIMATE-FEVER}: A Dataset for Verification of Real-World Climate Claims},
  author        = {Diggelmann, Thomas and Boyd-Graber, Jordan and Bulian, Jannis and Ciaramita, Massimiliano and Leippold, Markus},
  year          = {2020},
  eprint        = {2012.00614},
  archivePrefix = {arXiv}
}

@inproceedings{hasibi2017dbpedia,
  title     = {{DBpedia-Entity} v2: A Test Collection for Entity Search},
  author    = {Hasibi, Faegheh and Nikolaev, Fedor and Xiong, Chenyan and Balog, Krisztian and Bratsberg, Svein Erik and Kotov, Alexander and Callan, Jamie},
  booktitle = {Proceedings of the 40th International ACM SIGIR Conference on Research and Development in Information Retrieval},
  year      = {2017}
}

@inproceedings{thorne2018fever,
  title     = {{FEVER}: a Large-scale Dataset for Fact Extraction and {VER}ification},
  author    = {Thorne, James and Vlachos, Andreas and Christodoulopoulos, Christos and Mittal, Arpit},
  booktitle = {Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics},
  year      = {2018}
}

@inproceedings{boteva2016nfcorpus,
  title     = {A Full-Text Learning to Rank Dataset for Medical Information Retrieval},
  author    = {Boteva, Vera and Gholipour, Demian and Sokolov, Artem and Riezler, Stefan},
  booktitle = {Advances in Information Retrieval (ECIR 2016)},
  year      = {2016}
}

@misc{quora2017,
  title        = {First Quora Dataset Release: Question Pairs},
  author       = {Iyer, Shankar and Dandekar, Nikhil and Csernai, Korn{\'e}l},
  howpublished = {\url{https://quoradata.quora.com/First-Quora-Dataset-Release-Question-Pairs}},
  year         = {2017}
}

@inproceedings{cohan2020specter,
  title     = {{SPECTER}: Document-level Representation Learning using Citation-informed Transformers},
  author    = {Cohan, Arman and Feldman, Sergey and Beltagy, Iz and Downey, Doug and Weld, Daniel S.},
  booktitle = {Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics},
  year      = {2020}
}

@inproceedings{wadden2020scifact,
  title     = {Fact or Fiction: Verifying Scientific Claims},
  author    = {Wadden, David and Lin, Shanchuan and Lo, Kyle and Wang, Lucy Lu and van Zuylen, Madeleine and Cohan, Arman and Hajishirzi, Hannaneh},
  booktitle = {Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing},
  year      = {2020}
}