Dataset Card for TreeOfLife-200M

If you are looking for the original release TreeOfLife-200M dataset, as used in training BioCLIP 2 and presented the paper, please see Revision a8f38b4. The dataset, as presented here, was used to train BioCLIP 2.5 Huge; it completes the dataset cleaning process and resolves an issue where Observation.org occurrences were not included in the training data.

With 233 million images representing 933,798 taxa across the tree of life, TreeOfLife-200M is the largest and most diverse public ML-ready dataset for computer vision models in biology at release. This dataset combines images and metadata from four core biodiversity data providers: Global Biodiversity Information Facility (GBIF), Encyclopedia of Life (EOL), BIOSCAN-5M, and FathomNet to more than double the number of unique taxa covered by TreeOfLife-10M, adding 70 million more images than BioTrove (and nearly triple the unique taxa). TreeOfLife-200M also increases image context diversity with museum specimen, camera trap, and citizen science images well-represented (see Image Types). Our rigorous curation process ensures each image has the most specific taxonomic label possible and that the overall dataset provides a well-rounded foundation for training BioCLIP 2, BioCLIP 2.5 Huge, and future biology foundation models.

Dataset Details

Supported Tasks and Leaderboards

Image Classification, Zero-shot and few-shot Classification.

Dataset Contents

/dataset/
    catalog.parquet
    embeddings/
        README.md
        txt_emb_bioclip-2.json
        txt_emb_bioclip-2.npy
        txt_emb_bioclip-2.5-vith14.json
        txt_emb_bioclip-2.5-vith14.npy
        # duplicate of txt_emb_bioclip-2 files to support pybioclip backwards compatibility
        txt_emb_species.json
        txt_emb_species.npy
    metadata/
        Darwin-finches.csv
        eol_metadata/
            media_manifest.csv
            taxon.tab
        fathom_metadata.csv
        provenance.parquet   # For full Dataset
        resolved_taxa/       # product of TaxonoPy
            source=bioscan/
            source=eol/
            source=fathomnet/
            source=gbif/

BioCLIP 2 and BioCLIP 2.5 Huge text embeddings for all TreeOfLife-200M images used to train them are provided under embeddings/txt_emb_<model-name>.npy with associated labels embeddings/txt_emb_<model-name>.json. The BioCLIP 2 text embeddings and labels are duplicated as txt_emb_species.npy and txt_emb_species.json to maintain pybioclip functionality for versions 2.1.x and earlier. See embeddings/README.md for more information.

To avoid republishing existing datasets or interfering with original data source providers' ability to track use of their data, we provide all metadata here with step-by-step reproduction instructions in the GitHub repository to download the images and recreate the proper webdataset structure. This process will produce a collection of files named shard-######.tar in a train folder with which to work.

Inside each shard is a collection of images (named <uuid>.jpg), for which each has the following files:

<uuid>.com.txt
<uuid>.common_name.txt
<uuid>.jpg
<uuid>.sci.txt
<uuid>.sci_com.txt
<uuid>.scientific_name.txt
<uuid>.taxon.txt
<uuid>.taxonTag.txt
<uuid>.taxonTag_com.txt
<uuid>.taxon_com.txt
<uuid>.taxonomic_name.txt

Data Instances

Each image in this dataset is matched to the 7-rank Linnean taxonomy and common name of the subject of the image (where available). Examples of these text types is provided below. 89.74% (209,134,719) of the images have full taxonomic labels (for more context, see discussion on labeling challenges in biodiversity data under Considerations for Use). In addition to the biodiversity introduced (933,798 unique taxa), these images also span a variety of settings (or "images types"), the three main categories we highlight being museum specimen, camera trap, and citizen science images. Counts for these and more specific museum specimen subcategories (as used in processing) are provide in a table below.

Text Types

Text Type	Example
Common name	Narrow-leaved meadow-grass
Scientific name	Poa angustifolia
Taxonomic label	Plantae Tracheophyta Liliopsida Poales Poaceae Poa angustifolia

Image Types (GBIF only)

Image Type	Number of Images
Camera-trap	595.6K
Citizen Science	171.1M
Museum Specimen: Fungi	599.5K
Museum Specimen: Insect	6.9M
Museum Specimen: Invertebrate Zoology	1.7M
Museum Specimen: Microbiology	38.7K
Museum Specimen: Plant	39.7M
Museum Specimen: Uncategorized	1M
Museum Specimen: Vertebrate Zoology - Amphibians	36.2K
Museum Specimen: Vertebrate Zoology - Birds	345.7K
Museum Specimen: Vertebrate Zoology - Fishes	199.7K
Museum Specimen: Vertebrate Zoology - Mammals	129.1K
Museum Specimen: Vertebrate Zoology - Others	137K

Data Fields

catalog.parquet

• uuid: unique identifier for the image in this dataset, created at time of download.
• source_url: URL from which the image was downloaded, used to preview image in viewer.
• kingdom: kingdom to which the subject of the image belongs (Animalia, Plantae, Fungi, Chromista, Protozoa, Bacteria, Viruses, Archaea, and less than 1% unresolved cases, e.g., Metazoa and incertae sedis).
• phylum: phylum to which the subject of the image belongs.
• class: class to which the subject of the image belongs.
• order: order to which the subject of the image belongs.
• family: family to which the subject of the image belongs.
• genus: genus to which the subject of the image belongs.
• species: species to which the subject of the image belongs; note that this is the specific epithet.
• scientific_name: <Genus> <species> of the subject of the image, or most specific taxonomic rank available.
• common: vernacular name associated with the subject of the image where available. Common name is sourced from GBIF vernacular names provided in the GBIF Backbone Taxonomy.
• data_source: source from which the record was obtained (gbif, eol, bioscan, or fathomnet).
• publisher: the organization or entity that published the data (exclusive to GBIF records).
• basisOfRecord: basis of the record (e.g., PRESERVED_SPECIMEN, HUMAN_OBSERVATION, MACHINE_OBSERVATION; exclusive to GBIF records).
• img_type: the type of the associated image or its subject (e.g., Museum Specimen: Fungi, Citizen Science, Camera-trap), as defined by our processing (exclusive to GBIF records; EOL, BIOSCAN, and FathomNet images are all classified as `Unidentified).
• source_id: unique identifier for the image used by the source.
  • Note: for GBIF, this is gbifID which is unique to the occurrence containing the image. For EOL, this is <EOL Content ID>_<EOL Page ID>, the first of which is not a persistent identifier.
• shard_filename: name of shard containing the image.
• shard_file_path: filepath of the shard containing the image.
  • ex: 200M_v1.1/224x224/shards/shard-00123.tar
• base_dataset_file_path: filepath of the parquet file in the dataset containing the image.
  • ex: data/source=gbif/server=observation.org/data_7c3169bf-9670-4c03-86e7-c6d31436bb33.parquet

provenance.parquet: Provenance metadata for the full dataset.

• uuid: unique identifier for the image in this dataset, created at time of download.
• source_id: unique identifier for the image used by the source.
  • Note: for GBIF, this is gbifID which is unique to the occurrence containing the image. For EOL, this is <EOL Content ID>_<EOL Page ID>, the first of which is not a persistent identifier.
• data_source: source from which the record was obtained (gbif, eol, bioscan, or fathomnet).
• source_url: URL from which image was downloaded.
• license_name: name of license attached to the image (eg., cc-by).
• copyright_owner: copyright holder for the image, filled with not provided if no copyright owner was provided.
• license_link: URL to the listed license, left null in the case that license_name is No known copyright restrictions.
• title: title provided for the image, filled with not provided if no title was provided.
• bibliographicCitation: bibliographic citation for the occurrence record as provided by GBIF publishers. Only available for GBIF records where the publisher supplied a citation (~2.25% of records). Non-GBIF records (EOL, BIOSCAN, FathomNet) have null values.

Darwin-finches.csv: File for Darwin's finches embedding space evaluation completed in paper. Images are a represenative subset of the 18 species known as "Darwin's Finches" sampled from TreeOfLice-200M for this evaluation. Common names are from Avibase.

• uuid: unique identifier for the image in this dataset, links to other TreeOfLife-200M metadata.

• scientific_name: scientific name of the species.

• filepath: <scientific_name>/<uuid>.jpg the expected filepath to the images when calling them from a directory with the images organized in subfolders determined by the scientific name of the bird in the image.

• common_name: common name used by Avibase for the species.

• group: the "group" to which the bird belongs ('tree finches', 'warbler finches', 'ground finches', 'cactus finches', 'vampire finch', 'Cocos finch', or 'vegetarian finch'), as determined by name and the paper, Lamichhaney, S., Berglund, J., Almén, M. et al. Evolution of Darwin’s finches and their beaks revealed by genome sequencing. Nature 518, 371–375 (2015). https://doi.org/10.1038/nature14181.

  group image counts:

  group name	number of images
  ground finches	180
  tree finches	145
  cactus finches	133
  warbler finches	100
  Cocos finch	50
  vegetarian finch	50
  vampire finch	19

EOL-specific Files

media_manifest.csv: File with license, source, and copyright holder associated to each image from EOL listed in catalog.parquet and provenance.parquet; use <EOL Content ID>_<EOL Page ID> to match on source_id for provenance.parquet. Remaining columns are

• EOL content ID: unique identifier within EOL database for images sourced from EOL. Note that EOL content IDs are not stable.
• EOL page ID: identifier of page from which images from EOL are sourced. Note that an image's association to a particular page ID may change with updates to the EOL (or image provider's) hierarchy. However, EOL taxon page IDs are stable.
• Medium Source URL: URL pointing to source of image.
• EOL Full-Size Copy URL: URL to access the full-sized image; this is the URL from which the image was downloaded for this dataset (see Initial Data Collection and Normalization for more information on this process).
• License Name: name of license attached to the image (e.g., cc-by).
• Copyright Owner: copyright holder for the image, filled with not provided if no copyright owner was provided.

taxon.tab: Tab-delimited file with taxonomic information for EOL images based on EOL page IDs.

• taxonID: unique identifier for the file.
• source: often <source>:<id> where the source corresponds to the domain of the furtherInformationURL. The ID likely corresponds to an ID at the source.
• furtherInformationURL: URL with more information on the indicated taxon.
• acceptedNameUsageID: taxonID for the name accepted to represent this entry. Less than a third of these are non-null
• parentNameUsageID: taxonID of taxonomic rank above the indicated taxonRank in the hierarchy (eg., the taxonID of the genus Atadinus for the Atadinus fallax (Boiss.) Hauenschild entry).
• scientificName: scientific name associated with the EOL page (<canonicalName> <authority>, authority as available).
• taxonRank: lowest rank of the taxonomic tree indicated (eg., genus or species), occasionally not indicated, even for accepted names.
• taxonomicStatus: whether the name is accepted by EOL or not (accepted or not accepted, correspond to existence of non-null eolID or acceptedNameUsageID entry, respectively).
• datasetID: generally corresponds to the source identified in source column.
• canonicalName: the name(s) associate with the taxonRank (eg., <Genus> <species> for species).
• authority: usually name of person who assigned the name, with the year as available.
• eolID: the EOL page ID (only non-null when taxonomicStatus is accepted by EOL).
• Landmark: numeric values, meaning unknown, mostly null.
• higherClassification: labeling in the EOL Dynamic Hierarchy above the taxonRank (eg., Life|Cellular Organisms|Eukaryota|Opisthokonta|Metazoa|Bilateria|Protostomia|Ecdysozoa|Arthropoda|Pancrustacea|Hexapoda|Insecta|Pterygota|Neoptera|Endopterygota|Coleoptera|Adephaga|Carabidae|Paussus).

FathomNet Specific File

fathom_metadata.csv: Metadata file with FathomNet images identifiers and associated bounding boxes. This is required to fetch FathomNet images since there are multiple bounding boxes used from some images to generate final TreeOfLife-200M images.

• uuid: unique identifier to match to catalog and provenance parquet files.
• x,y: together these columns give the reference coordinate from which to construct the bounding boxes.
• width: width in pixels of bounding box the image came from.
• height: height in pixels of bounding box the image came from. This file is required along with provenance.parquet to fetch all the FathomNet images since some are pulled from bounding boxes around different creatures in the same base image.

Data Splits

This entire dataset was used for training the model, BioCLIP 2.5 Huge (similarly, all of Revision a8f38b4 was used to train BioCLIP 2). We used 11 biologically-relevant datasets for various species classification tests of the model trained on this dataset; they are described (briefly) and linked to below.

Test Sets

BioCLIP 2 and BioCLIP 2.5 Huge were tested on the same 10 biologically-relevant benchmarks as BioCLIP, though we used NABirds in place of Birds 525, since the latter is no longer available online. We also curated IDLE-OO Camera Traps.

• NABirds: 48K images of the 400 most common North American birds (based on observation counts), with at least 100 images per species. Includes age and sex annotations that we used for evaluation.
• Meta-Album: Specifically, we used the Plankton, Insects, Insects 2, PlantNet, Fungi, PlantVillage, and Medicinal Leaf datasets.
• Rare Species: Balanced benchmark dataset with 400 species (30 images per species) listed under 5 IUCN Red List categories (Near Threatened, Vulnerable, Endangered, Critically Endangered, and Extinct in the Wild).
• IDLE-OO Camera Traps: A new benchmark we curated from a combination of 5 different LILA BC camera trap datasets with image-level labels. These were filtered to contain only images with a single species of animal in frame.

Additional hierarchical structure and embedding space evaluation was done using images of Darwin's Finches sourced from the training data (in metadata/Darwin-finches.csv). These species are not evenly represented in the dataset, so 12 of them have 50 images each, but the remaining six species (Geospiza propinqua, Geospiza acutirostris, Camarhynchus psittacula, Geospiza septentrionalis, Camarhynchus pauper, Camarhynchus heliobates) have between 8 and 33 representative images (listed in decreasing order).

Other non-species classification tasks were also used for testing, and are described in the BioCLIP 2 and BioCLIP 2.5 Huge model card and our paper.

Dataset Creation

Curation Rationale

TreeOfLife-200M was curated for the purpose of training a biological foundation model. In particular, we aimed to increase the both the biodiversity of available training data (i.e., from that available in TreeOfLife-10M) and the raw number of images available (i.e., from that of BioTrove). We also performed extensive data curation beyond that used for either of these training datasets: aligning taxonomic labels, removing images of labels or folders associated to specimens, removing noisy citizen science and camera trap images, and ensuring no identifiable humans are in the images (more info below).

We have expanded coverage across the 2.14M described species estimated by The International Union for Conservation of Nature (IUCN) as of March 2025, with particularly strong representation of threatened species (70.38% of species across threatened categories are represented in TreeOfLife-200M). This coverage and the implications are discussed in greater detail in our paper.

Source Data

TreeOfLife-200M is composed of images from the following four core data providers:

1. Global Biodiversity Information Facility (GBIF), which is a major biological data aggregator cataloging biodiversity data from citizen science sources (e.g., iNaturalist and Observation.org), museum collections (e.g., from the Smithsonian Institution, Museum national d'Histoire naturelle), and camera trap collections (e.g., from the Research Institute for Nature and Forest (INBO)).

• GBIF provides a DOI for the occurrence snapshot we downloaded (GBIF occurrence 2024-05-01 snapshot with filter "occurrenceStatus": "PRESENT" (DOI: https://doi.org/10.15468/dl.bfv433)).
• This manifest was further filtered for only Still Images, those with gbifID and identifier, and those not labeled as images of text documents prior to download.
  • Note: there is only one category ("basisOfRecord": "MATERIAL_CITATION") that describes images of textual documents (these are a particular type of document that describes species).
• Further filtereing and processing steps are described below.

2. Encyclopedia of Life (EOL): Another large biological data aggregator (greatest source of biodiversity in TreeOfLife-10M).

• EOL does not have a versioned release like GBIF, so recreating this dataset requires use of metadata/eol_metadata/media_manifest.csv and metadata/eol_metadata/taxon.tab (this is just if one wants to reproduce our process, the provided catalog and provenance parquets are sufficient to reproduce a copy of the dataset).
  • Media manifest was downloaded from "image list URL: https://eol.org/data/media_manifest.tgz" (at image list). Taxonomy (taxon.tab) from "EOL Dynamic Hierarchy Active Version URL: https://editors.eol.org/uploaded_resources/00a/db4/dh21.zip" (at dataset).
• Images used for the Rare Species dataset were removed from this collection (by MD5) to avoid data leakage.
• Further filtering and processing steps are described below.

3. BIOSCAN-5M: Collection of primarily insect specimen images, hand-labeled by experts. Other Arthropoda classes account for 2% of the 5M images. 93% of the images in the dataset are not labeled to the species level.

• This dataset was ready to use, and we followed access instructions in their repository (specifically downloading through the Google Drive link).
• The inclusion of this data just required aligning the taxon labels with those of the other datasets (following the Annotation Process, described below).

4. FathomNet: Collection of expert-annotated images of undersea creatures; this dataset focuses on a habitat instead of a clade.

• This dataset required fetching image metadata through their API to generate a manifest, then a second pass to fetch bounding boxes for each image.
• From these, "concepts" were extracted and GNVerifier was utilized to obtain the relevant taxonomic information.
  • Extracted taxonomic information required alignment (as described in Annotation Process, below).
• Focal species were retrieved from each image following download using distributed-downloader.
  • Images with multiple focal species had each extracted into its own image utilizing the bounding box annotations, hence the inclusion of a fathom_metadata CSV to fetch all images. The process is described, with specific instructions, in TreeOfLife-toolbox/docs/FathomNet_download_README.

Both BIOSCAN-5M and FathomNet were included to help improve coverage of under-represented and highly diverse branches of the tree of life (Insecta is one of the most diverse classes and creatures that dwell in the ocean are much less commonly represented). The total number of dataset-wide taxonomic hierarchies that are uniquely contributed by each core data provider is provided below, with their total number of images contributed to exemplify this point.

Provider	Unique Taxa	Images
GBIF	569,225	222.6M
EOL	48,508	5.2M
BIOSCAN-5M	3,042	5.2M
FathomNet	251	37.9K

Data Curation and Processing

During download (with distributed-downloader), each image's size is checked: first that it is sufficiently large (224 pixels or more on shortest side) and excluded if not, and then the image is resized, if necessary, so that its largest side does not exceed 720 pixels. A hash of both the original and the resized image is recorded, and the hash of the original image is compared across the dataset to avoid introducing duplicate images. The metadata is still recorded, so that no information is lost (images may be uploaded to one site with more or less information than to another).

The taxonmic alignment process, which was applied to all source data, is described below in Annotations. The remaining processing was focused on (1) ensuring image quality and (2) eliminating duplication and data leakage, and is described below.

1. Ensuring Image Quality

We noted the varied image types included in the GBIF snapshot, above. The three main types (musuem specimen, camera trap, and citizen science images) each require special filtering processes to avoid introducing noisy data.

Musuem specimen collections are often photographed with both specimens and their associated text documents (e.g., labels, folders, tags) under the same occurrence. All images under a single occurrence share metadata, there is no distinction for the subject of the image. These images of text were filtered through a nearest-centroid classifier spanning 25 fine-grained subtypes; 11 collection areas (Fungi, Insect, Invertebrate Zoology, Microbiology, Plant, Uncategorized, and five classes of Vertebrate Zoology: Amphibians, Birds, Fishes, Mammals, Others) inspired by the Smithsonian Institution's collection categories, each further subdivided to classes such as preserved or fossil specimen, as labed by GBIF basisOfRecord. We manually curate an 8.5K set of examples on which to fit the classifier, which predicts a subtype for all museum images. All non-organismal images are then removed.

Camera trap occurrences are first filtered to only those with 15 images or less. MegaDetector is then applied to each occurrence to distinguish those with animals in them from "empty" frames. We remove these to avoid labeling a creature's habitat as the creature, particularly since we are training a foundation model for the full tree of life, meaning that plants are of interest in their own right and should not be mislabeled as animals.

• Context: Some occurrences include large-volume camera-trap sequences, with up to 10,000 images. These occurrences have a single taxonomic label applied across all images, though there are different taxa (e.g, the first few frames had a duck, while later images have a swan, another a goose, but the label for all is a duck). In these cases, Megadetector often misclassifies empty frames or detects animals that do not match the intended taxa, leading to false positives. To reduce the risk of introducing such noise while still capturing relevant biodiversity, we filter the dataset to include only occurrences with 15 images or fewer.

Finally, images within citizen science occurrences are tested for similarity, and the mean pair-wise BioCLIP embedding distance is used to determine whether they fall into one of three categories: (1) those that are overly distinct (this can happen when images of different species are uploaded to the same observation), (2) those that exhibit "expected" variation, and (3) those that are exceedingly similar (this can occur when images from a single camera trap are uploaded to one observation). The images in (1) are removed (bottom 5th percentile), those in (2) are retained, and those in (3) are run through the camera trap processing described above.

We run MTCNN on all images from GBIF and EOL to detect and remove images containing identifiable human faces. Approximately 10K fish images determined to have human faces were recovered in the 233M image update by cropping the images to just the fish. BIOSCAN-5M and FathomNet Database do not have images with human faces requiring filtering.

2. Eliminating Duplication and Data Leakage

As noted above, MD5 hashes are taken of every image on download. This are highly sensitive, in that variation by a single pixel will result in a different hash between two images. Thus, these are used for an initial filter across all images to avoid duplication and the introduction of test set images (e.g., those from Rare Species that were sourced from EOL) into the training data. The next step is to check for test images by metadata; however, this is not always viable, as existing benchmarks do not always retain the connection to their orignal source data in a manner allowing this. We thus take perceptual PDQ hashes of every test image and all images from GBIF and EOL (with Medium Source URL from Flickr--these methods are not effective on museum specimen images), removing those with distance less than 10 from the training data.

The code for all of these processing steps, along with further details, is provided in TreeOfLife-toolbox.

Annotations

We standardized the taxonomic labels provided by the four core data providers to conform to a uniform 7-rank Linnean structure.

Kingdom Counts

Kingdom	Number of Images
Animalia	114904588
Plantae	107063962
Fungi	9154557
Chromista	530074
Protozoa	214751
Bacteria	59181
Viruses	8452

Annotation process

Taxonomic labels (kingdom, phylum, etc.) were standardized across the various data sources using the TaxonoPy package v0.1.0 that we designed (in consultation with taxonomists) for this purpose. The TaxonoPy algorithm works to match each unique taxonomic string to the GBIF Backbone, Catalogue of Life, and OpenTree hierarchies (in that order). See TaxonoPy for more details on this process.

Who are the annotators?

The taxonomic labels provided by the four core data providers (GBIF, EOL, BIOSCAN-5M, and FathomNet) were fed into TaxonoPy for standardization.

Personal and Sensitive Information

None: These images come from existing, public biodiversity data repositories, and no GPS locations for the species in the images are included. We further used MTCNN on all images downloaded from GBIF and EOL to filter out identifiable images of people. There are no images with people in the BIOSCAN-5M or FathomNet Database to filter.

Considerations for Using the Data

We hope this dataset (and the model trained with it) can aid in conservation efforts and biodiversity research.

Bias, Risks, and Limitations

This dataset is imbalanced in its representation of various species with the greatest representation available for those in the phyla Tracheophyta, Arthropoda, and Chordata. The long-tailed distribution exhibited by taxonomic rank is to be expected working with biodiversity data, and results from both an imbalance in the availability of images and actual variance in class diversity. For instance, TreeOfLife-200M has a balanced representation between plants and animals (at the kingdom level), but animals represent a larger proportion of described species. Additionally, IUCN estimates 2,139,242 described species as of March 2025, the vast majority of which are invertebrates. Though described invertebrates outnumber vertebrates 20 to one, more than double the number of vertebrates have been evaluated (64K vertebrates to only 29K invertebrates) due to lack of available information.

Overall, 89.74% of images have full taxonomic labels. As discussed above, most of this gap is more indicative of the lack of consensus or available granularity in labeling, rather than missing information. BIOSCAN-5M is a good example of this, as the label granularity is still limited for insects (Insecta, one of the most diverse classes in the tree of life). In our resolved taxa for BIOSCAN-5M, 95.97% of images are labeled to the family level but only 30.74% and 6.81% of the images have genus or species indicated, respectively. This is a persistent challenge in species identification and part of the reason that large biological foundation models trained with taxonomic hierachies are useful. We highlight these taxonomic coverage limitations for a better understanding of both the data and our motivation, but do not wish to diminish the impact of our coverage of nearly 868K unique taxa labeled to the level of species; it far surpasses the species diversity of other well-known biological datasets.

We note also that our taxonomic resolution (TaxonoPy) resulted in less than 1% unresolved cases, e.g., Metazoa instead of Animalia or incertae sedis for kingdom:

Unresolved Kingdom	Number of Images
incertae sedis	1056775
Archaeplastida	33891
Metazoa	24461
...
NULL	34
...

Recommendations

It is always important to understand the data with which one is working and take each dataset's uniqueness into account. Some considerations for TreeOfLife-200M:

• If using this dataset as more than just a training dataset (i.e., for validation or testing) be sure to keep source_id distinct across splits.
  • In these circumstances, it would also good to keep in mind that this dataset has a long-tail distribution (as one would expect from biological data).
• As noted above, there are some unresolved/misaligned kingdoms. Depending on one's use-case, these may be dropped or another attempt at re-alignment might be made (for instance with an updated version of TaxonoPy, as we intend to add profiles to address these outliers).

Licensing Information

The data (images and text) contain a variety of licensing restrictions mostly within the CC family. Each image and text in this dataset is provided under the least restrictive terms allowed by its licensing requirements as provided to us (i.e, we impose no additional restrictions past those specified by licenses in the license file).

All BIOSCAN-5M images are licensed under Creative Commons Attribution 3.0 Unported (CC BY 3.0).

FathomNet: All FathomNet images and annotations are shared under Creative Commons Attribution-Non Commercial-No Derivatives 4.0 International license (CC BY-NC-ND, more details on their terms).

EOL and GBIF images contain a variety of licenses ranging from CC0 to CC BY-NC-SA. The GBIF occurrence snapshot is licensed under Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC)

For license and citation information by image, see our provenance file.

This dataset (the compilation) has been marked as dedicated to the public domain by applying the CC0 Public Domain Waiver. However, images are often licensed under different terms (as noted above).

Citation

If you use this dataset in your research, please cite both it and our associated paper, as well as the constituent datasets (source data providers).

BibTeX:

Data

@dataset{treeoflife_200m,
  title = {{T}ree{O}f{L}ife-200{M} (Revision <commit>)}, 
  author = {Jianyang Gu and Samuel Stevens and Elizabeth G Campolongo and Matthew J Thompson and Net Zhang and Jiaman Wu and Andrei Kopanev and Zheda Mai and Alexander E. White and James Balhoff and Wasila M Dahdul and Daniel Rubenstein and Hilmar Lapp and Tanya Berger-Wolf and Wei-Lun Chao and Yu Su},
  year = {2026},
  url = {https://huggingface.co/datasets/imageomics/TreeOfLife-200M},
  doi = {<update on release>},
  publisher = {Hugging Face}
}

Note that this version is updated from Revision a8f38b4, which was used to train the BioCLIP 2 model, and was presented in the paper. This updated version completes the dataset cleaning process and resolves an issue where Observation.org occurrences were not included in the training data. It was used to train BioCLIP 2.5 Huge. Approximately 2.9 million images from Revision a8f38b4 have been filtered out, but an additional 21.8M images were added (some recovered through targeted segmentation, most from inclusion of Observation.org images). The BioCLIP 2.5 Huge text embeddings were also added.

Please also cite our paper:

@inproceedings{NEURIPS2025_94da80cb,
 author = {Gu, Jianyang and Stevens, Sam and Campolongo, Elizabeth and Thompson, Matthew and Zhang, Net and Wu, Jiaman and Kopanev, Andrei and Mai, Zheda and White, Alexander and Balhoff, James and Dahdul, Wasila and Rubenstein, Daniel and Lapp, Hilmar and Berger-Wolf, Tanya and Chao, Wei-Lun (Harry) and Su, Yu},
 booktitle = {Advances in Neural Information Processing Systems},
 editor = {D. Belgrave and C. Zhang and H. Lin and R. Pascanu and P. Koniusz and M. Ghassemi and N. Chen},
 pages = {102778--102811},
 publisher = {Curran Associates, Inc.},
 title = {BioCLIP 2: Emergent Properties from Scaling Hierarchical Contrastive Learning},
 url = {https://proceedings.neurips.cc/paper_files/paper/2025/file/94da80cbfe870c1db958c88a8a27018c-Paper-Conference.pdf},
 volume = {38},
 year = {2025}
}

Please be sure to also cite the original data sources and all constituent parts as appropriate.

Source Data Providers

GBIF:

@misc{GBIF-DOI,
  doi = {10.15468/DL.BFV433},
  url = {https://doi.org/10.15468/dl.bfv433},
  keywords = {GBIF, biodiversity, species occurrences},
  author = {GBIF.org},
  title = {{GBIF} Occurrence Download},
  publisher = {The Global Biodiversity Information Facility},
  month = {May},
  year = {2024},
  copyright = {Creative Commons Attribution Non Commercial 4.0 International}
}

EOL:

@misc{eol,
    author = {{Encyclopedia of Life (EOL)}},
    url = {https://eol.org},
    note = {Accessed August 2024}
}

BIOSCAN-5M:

@inproceedings{gharaee2024bioscan5m,
    title={{BIOSCAN-5M}: A Multimodal Dataset for Insect Biodiversity},
    booktitle={Advances in Neural Information Processing Systems},
    author={Zahra Gharaee and Scott C. Lowe and ZeMing Gong and Pablo Millan Arias
        and Nicholas Pellegrino and Austin T. Wang and Joakim Bruslund Haurum
        and Iuliia Zarubiieva and Lila Kari and Dirk Steinke and Graham W. Taylor
        and Paul Fieguth and Angel X. Chang
    },
    editor={A. Globerson and L. Mackey and D. Belgrave and A. Fan and U. Paquet and J. Tomczak and C. Zhang},
    pages={36285--36313},
    publisher={Curran Associates, Inc.},
    year={2024},
    volume={37},
    url={https://proceedings.neurips.cc/paper_files/paper/2024/file/3fdbb472813041c9ecef04c20c2b1e5a-Paper-Datasets_and_Benchmarks_Track.pdf},
}

FathomNet:

@article{katija_fathomnet_2022,
    title = {{FathomNet}: {A} global image database for enabling artificial intelligence in the ocean},
    volume = {12},
    issn = {2045-2322},
    shorttitle = {{FathomNet}},
    url = {https://www.nature.com/articles/s41598-022-19939-2},
    doi = {10.1038/s41598-022-19939-2},
    number = {1},
    urldate = {2025-05-13},
    journal = {Scientific Reports},
    author = {Katija, Kakani and Orenstein, Eric and Schlining, Brian and Lundsten, Lonny and Barnard, Kevin and Sainz, Giovanna and Boulais, Oceane and Cromwell, Megan and Butler, Erin and Woodward, Benjamin and Bell, Katherine L. C.},
    month = sep,
    year = {2022},
    pages = {15914},
}

For license and citation information by image, see our provenance file.

Acknowledgements

We would like to thank Zhiyuan Tao, Shuheng Wang, Ziheng Zhang, Zhongwei Wang, and Leanna House for their help with the TreeOfLife-200M dataset, Charles (Chuck) Stewart, Sara Beery, and other Imageomics Team members for their constructive feedback and Sergiu Sanielevici, Tom Maiden, and TJ Olesky for their dedicated assistance with arranging the necessary computational resources.

We are grateful to Kakani Katija and Dirk Steinke for helpful conversations regarding use and integration of FathomNet and BIOSCAN-5M, respectively, as well as Stephen Formel and Markus Döring for GBIF. We thank Marie Grosjean for comparative methods for filtering citizen science images and Dylan Verheul for assistance with acquiring images from Observation.org from GBIF. We thank Suren Byna for a helpful conversation on early dataset design decisions. We thank Doug Johnson for his collaboration in hosting this large dataset on the Ohio Supercomputer Center research storage file system.

This work was supported by the Imageomics Institute, which is funded by the US National Science Foundation's Harnessing the Data Revolution (HDR) program under Award #2118240 (Imageomics: A New Frontier of Biological Information Powered by Knowledge-Guided Machine Learning).

Our research is also supported by resources from the Ohio Supercomputer Center.

Any opinions, findings and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the National Science Foundation.

Dataset Card Authors

Elizabeth G. Campolongo