PHVL: audited and unified public plant disease, pest and weed image datasets for vision-language models

Metadata, copy groups, labels and a host–condition ontology (PHVL) that unify 28 public plant disease, pest and weed image datasets for training vision-language models. No images are included: the per-image manifest points to the original downloads, which keep their own licences. phvl_image_manifest.tsv.gz: 1,250,418 audited images with difference hash, copy group (identical, near-duplicate and rotated/flipped copies confirmed by pixel correlation), canonical flag (783,893 unique images), evaluation-benchmark contamination, training-base flag (469,789 images), original label and ontology concept. phvl.ttl: the PHVL ontology in SKOS (725 host–condition concepts over 70 hosts; hosts and 368 causal agents checked against NCBI Taxonomy and linked to EPPO, AGROVOC and Wikidata; links to the Plant Ontology and the Plant Stress Ontology), with tables phvl_concepts.csv and phvl_class_mapping.csv. Dataset cards (origin, lineage, class inventory, label semantics, capture, splits, quality, licence), audit tables (lineage graph, canonical sources, licence-inheritance conflicts, label conflicts, label agreement between copies) and the code. The condition curation was assisted by large language models and checked against NCBI Taxonomy; low-confidence concepts are flagged and await expert review. See README.md. Funding: São Paulo Research Foundation (FAPESP), grant #2026/18049-6.