An expert-curated sentence-level corpus of lipid–phenotype and lipid–function relations, with release tiers

An expert-curated, sentence-level corpus of lipid–phenotype and lipid–biological-function relation statements extracted from PubMed titles and abstracts. The corpus comprises 7,524 sentence-level relation records (7,489 distinct sentences) drawn from 7,464 distinct PubMed articles, divided into 4,060 lipid–phenotype (PF) and 3,464 lipid–biological-function (BF) records. Each record provides the normalized lipid and target entities with 1-indexed inclusive character offsets, the relation cue, and four independent expert labels (YES / No / unknown) recorded in a single pass with no discussion and no reconciliation step. A deterministic rule assigns every record to one of four release tiers: positive agreement (T1, 6,559), negative agreement (T2, 279), mixed disagreement (T3, 309), and unresolved (T4, 377). The negative-agreement and mixed-disagreement tiers are released alongside the positives, so the corpus provides expert-vetted negative controls and a set of hard, genuinely contested cases rather than positives alone. A PMID-grouped, stratified train/development/test split (4,789 / 1,142 / 1,216) is provided for tiers T1–T3. Inter-annotator agreement on the combined pool: Fleiss κ = 0.629, Krippendorff α = 0.629, percent agreement = 92.5%, Gwet AC1 = 0.916. The deposit includes the main corpus and its two sub-corpora, the split assignments, a Tier-3 disagreement taxonomy, the annotation guideline used by the annotators, a full data dictionary, an automated quality-control summary (0 hard failures), and the analysis code. Annotator identities are anonymized throughout. This dataset accompanies a Scientific Data Data Descriptor. The curation pool originates from the LipidFun database (Lin et al., Bioinformatics 2025, doi:10.1093/bioinformatics/btaf110), which did not release the annotated sentences.