Pollen AI Atlas: million-scale multimodal pollen microscopy with expert-guided foundation models

This Zenodo release is the public processed-data package accompanying the Pollen AI Atlas paper, "Million-scale multimodal pollen microscopy with expert-guided foundation models". The package contains filtered detection metadata for all 85 processed whole-slide microscopy images; five caption sets for the 80-slide caption subset; 1,511,390 pre-extracted pollen-grain image crops from pure-species slides; 17,318 expert-validated ground-truth pollen crops; train/validation split definitions; supporting slide-level metadata; caption anchors; query images; and the fine-tuned LVD-ViT-S checkpoint used in downstream processing. The caption files include Gemma4 as the primary caption set, with Qwen2.5-VL, Qwen3-VL, Qwen3.5, and Qwen3.6 provided as companion caption corpora. Caption records are provided as per-slide JSONL files. Filtered detections are provided as HDF5 files. Expert annotations are provided as GeoJSON files. Image crops are provided as PNG files, with CSV manifests linking crop filenames to grain identifiers and whole-slide coordinates. The release covers 1,511,390 public pollen-grain objects spanning four dataset origins: French, Hungarian, Mediterranean, and Swedish. Five slides are excluded from the caption subset for taxonomy or quality reasons, but their filtered detection metadata are included where applicable. The 85 original whole-slide microscopy images are not included in this Zenodo package and remain available from the corresponding authors upon reasonable request. Please cite both the Zenodo record and the associated manuscript.