TreeOfLife-200M Biodiversity Image Dataset

## Overview TreeOfLife-200M is the largest public ML-ready biodiversity image dataset at release, containing approximately 214 million images representing 952,257 taxa across the tree of life. The dataset unifies images and taxonomic metadata from four major biodiversity sources into a single parquet-based collection optimized for computer vision and multimodal model training (e.g., CLIP-style species recognition). Languages in metadata include English and Latin (scientific names). ## Schema - `image` — binary/image — the specimen or observation photograph - `taxon` — string — full taxonomic identifier - `kingdom` / `phylum` / `class` / `order` / `family` / `genus` / `species` — string — taxonomic hierarchy fields - `scientific_name` — string — Latin binomial - `common_name` — string — vernacular name where available - `source` — string — provider (GBIF, EOL, BIOSCAN-5M, FathomNet) - `source_id` — string — original record identifier from provider - `license` — string — per-image upstream license tag - +several more metadata columns (provenance, geolocation where available) ## Sources - Hugging Face: https://huggingface.co/datasets/imageomics/TreeOfLife-200M — License: CC0-1.0 - Upstream: GBIF (Global Biodiversity Information Facility), Encyclopedia of Life (EOL), BIOSCAN-5M, FathomNet - Paper: arXiv:2505.23883 ## Methodology The Imageomics team aggregated images and taxonomic metadata from four core biodiversity data providers, harmonizing taxonomic labels against a unified tree-of-life hierarchy and deduplicating across sources. Images are sourced from citizen-science observations (GBIF/EOL), curated insect barcoding specimens (BIOSCAN-5M), and underwater/marine imagery (FathomNet). The dataset more than doubles the unique taxa coverage compared to prior biology CV datasets and is packaged as parquet for efficient streaming. ## Known gaps & limitations Class distribution is highly imbalanced (long-tail across taxa) — the dataset card explicitly flags this. Geographic and taxonomic sampling bias inherits from contributor activity in upstream providers (GBIF/EOL skew toward well-studied regions and charismatic species). Image quality varies widely across sources (lab specimens vs. in-situ field photos vs. marine ROV footage). Metadata completeness varies; not all records have full taxonomic hierarchies or geolocation. Buyers training on this data should validate label accuracy empirically and consider rebalancing strategies. ## Intended use & out-of-scope - Intended: training and evaluating biology-domain vision models, CLIP-style zero-shot species classification, fine-tuning foundation models for ecological/taxonomic tasks, knowledge-guided multimodal learning. - Out-of-scope: not deduplicated against common biology eval benchmarks (e.g., iNat21, BIOSCAN test splits) — leakage risk if used naively for benchmark training; not suitable for fine-grained individual-organism identification within a species. _Federated dataset: 1 parquet shards, 12.92 GB total. Queries and downloads stream through the DataBazaar API._ _PII signals: us_phone×9 present in the sample. Common in public datasets (papers, logs) but worth knowing before joining with private data._ Original supplier listing: TreeOfLife-200M: Biodiversity Image Dataset (214M images, 952K taxa) 214M biology/species images across 952,257 taxa from GBIF, EOL, BIOSCAN-5M, and FathomNet. CC0 licensed, parquet format, ready for CV/CLIP training and zero-shot taxonomic classification.