TreeOfLife-10M Biological Organism Images

## Overview TreeOfLife-10M is the largest-to-date ML-ready dataset of biological organism images paired with taxonomic labels, containing over 10 million images covering 454,000 taxa across the tree of life. Distributed in WebDataset format with image, text (taxonomic label), and document modalities. It expands on iNat21 and BIOSCAN-1M by incorporating newly curated images from the Encyclopedia of Life (eol.org). ## Schema - image — binary — JPEG/PNG image of a biological organism - taxonomic label — text — Linnaean taxonomy string (kingdom→species) - source — text — origin dataset (iNat21, BIOSCAN-1M, EOL) - taxon_id — text — identifier in source taxonomy - common_name — text (where available) — vernacular name - scientific_name — text — Latin binomial - Additional WebDataset shard metadata fields ## Sources - HuggingFace: https://huggingface.co/datasets/imageomics/TreeOfLife-10M - Underlying sources: iNaturalist 2021 (iNat21), BIOSCAN-1M, Encyclopedia of Life (eol.org) - License: per HuggingFace dataset card (Imageomics project, generally open for research and ML use; component datasets carry their own licenses, mostly CC-BY family) - DOI: 10.57967/hf/7542 ## Methodology The Imageomics Institute aggregated three large biological image sources: iNat21 (citizen-science observations from iNaturalist), BIOSCAN-1M (curated insect specimens with DNA barcodes), and newly curated Encyclopedia of Life imagery (which supplies the majority of taxa coverage). Images were paired with taxonomic labels harmonized against a unified tree-of-life taxonomy, packaged as WebDataset shards for efficient streaming during large-scale model training (e.g., BioCLIP). ## Known gaps & limitations Substantial class imbalance — the dataset is explicitly tagged "imbalanced" because real-world species observation frequency follows a long-tail distribution. iNat21-sourced images are biased toward charismatic and accessible species in well-surveyed regions (North America, Europe). Some taxa are represented by only a handful of images. Taxonomic labels reflect a snapshot of evolving biological consensus and may contain misidentifications inherited from citizen-science source data. Image quality and framing vary widely across source datasets. ## Intended use & out-of-scope - IS for: training vision-language models for biology (BioCLIP-style), zero-shot species classification, taxonomic retrieval, ecological CV research, fine-tuning species ID models. - NOT for: authoritative species identification in safety-critical contexts (e.g., toxic/edible determination), commercial use without checking component dataset licenses, training models on balanced taxonomic priors without reweighting. _Federated dataset: 10 parquet shards, 4.61 GB total. Queries and downloads stream through the DataBazaar API._ Original supplier listing: TreeOfLife-10M: 10M Biological Organism Images with Taxonomic Labels Largest ML-ready dataset of biological organism images (10M+ images, 454K taxa) paired with taxonomic labels. Aggregates iNat21, BIOSCAN-1M, and Encyclopedia of Life. Used to train BioCLIP and similar vision-language models.