denseon-pretrain-50m-balanced

50 million (query, document) pairs sampled from lightonai/embeddings-pre-training-curated across all 34 source subsets using balanced T=2 temperature sampling.

Sampling strategy

Pairs are sampled uniformly at random. Allocation per subset uses T=2 temperature weighting (alloc ∝ √source_size) with iterative equal redistribution: any dataset whose T=2 share exceeds its actual size is capped, and the surplus is split equally across the remaining datasets. This repeats until convergence, giving small datasets meaningful representation without repetition.

Allocation

Dataset	Pairs	% of total
mtp	6,994,442	13.99%
fw_edu	6,011,301	12.02%
reddit	4,612,121	9.22%
paq	3,534,179	7.07%
s2orc_title_abstract	3,005,283	6.01%
amazon_reviews	2,944,981	5.89%
s2orc_citation_titles	2,603,034	5.21%
s2orc_abstract_citation	2,553,298	5.11%
reddit_body_comment	1,962,252	3.92%
stackoverflow_title_body	1,819,604	3.64%
wikipedia_hlp_cm	1,609,716	3.22%
wikipedia_hlp_dl	1,609,716	3.22%
wikianswers	1,608,917	3.22%
stackexchange_title_body	1,092,099	2.18%
msmarco	994,320	1.99%
stackexchange_qa	969,315	1.94%
gooaq_qa	810,880	1.62%
beir_dbpedia	756,292	1.51%
arxiv_title_abstract	701,083	1.40%
yahoo_qa	478,936	0.96%
amazon_qa	453,071	0.91%
yahoo_answer	393,817	0.79%
agnews	391,470	0.78%
yahoo_question_body	350,207	0.70%
npr	343,729	0.69%
cc_news_en	281,762	0.56%
cnn_dailymail	281,762	0.56%
biorxiv_title_abstract	275,000	0.55%
medrxiv_title_abstract	189,000	0.38%
wikihow	128,000	0.26%
altlex	83,100	0.17%
stackexchange_duplicate_questions	56,800	0.11%
stackexchange_body_body	55,600	0.11%
quora	44,900	0.09%

Schema

Column	Type	Description
anchor	string	Query / question
positive	string	Relevant document / answer

Usage

from datasets import load_dataset

# All 50M pairs
ds = load_dataset("capemox/denseon-pretrain-50m-balanced", split="train")

# Single subset
ds = load_dataset("capemox/denseon-pretrain-50m-balanced", "msmarco", split="train")