Update Dataset Card

Dataset card drafted with @egrace479 , update information based on the
new TOL-200M catalog.

See PR 126 in GH repo for edit history.

README.md

@@ -17,6 +17,8 @@ tags:
 - image
 - imageomics
 - animals
+- plants
+- fungi
 - evolutionary biology
 - CV
 - multimodal
@@ -28,20 +30,23 @@ tags:
 - imbalanced
 size_categories:
 - 100M<n<1B
-description: "With nearly 214 million images representing 952,257 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 50 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. 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 and future biology foundation models."
+description: "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 50 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. 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 Card for TreeOfLife-200M
 
-With nearly 214 million images representing 952,257 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](https://gbif.org)), Encyclopedia of Life ([EOL](https://eol.org)), [BIOSCAN-5M](https://github.com/bioscan-ml/BIOSCAN-5M), and [FathomNet](https://www.fathomnet.org/) to more than double the number of unique taxa covered by [TreeOfLife-10M](https://huggingface.co/datasets/imageomics/TreeOfLife-10M), adding 50 million more images than [BioTrove](https://huggingface.co/datasets/BGLab/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](#image-types-gbif-only)). Our rigorous [curation process](#data-curation-and-processing) ensures each image has the most specific taxonomic label possible and that the overall dataset provides a well-rounded foundation for training [BioCLIP 2](https://huggingface.co/imageomics/bioclip-2) and future biology foundation models.
+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](https://huggingface.co/datasets/imageomics/TreeOfLife-200M/tree/a8f38b4388579862c56ae57d6f094c2ac0e92e12). The dataset, as presented here, was used to train [BioCLIP 2.5 Huge](https://huggingface.co/imageomics/bioclip-2.5-vith14); 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](https://gbif.org)), Encyclopedia of Life ([EOL](https://eol.org)), [BIOSCAN-5M](https://github.com/bioscan-ml/BIOSCAN-5M), and [FathomNet](https://www.fathomnet.org/) to more than double the number of unique taxa covered by [TreeOfLife-10M](https://huggingface.co/datasets/imageomics/TreeOfLife-10M), adding 70 million more images than [BioTrove](https://huggingface.co/datasets/BGLab/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](#image-types-gbif-only)). Our rigorous [curation process](#data-curation-and-processing) ensures each image has the most specific taxonomic label possible and that the overall dataset provides a well-rounded foundation for training [BioCLIP 2](https://huggingface.co/imageomics/bioclip-2), [BioCLIP 2.5 Huge](https://huggingface.co/imageomics/bioclip-2.5-vith14), and future biology foundation models.
 
 ## Dataset Details
 
 ### Dataset Description
 
 - **Curated by:** Jianyang Gu, Samuel Stevens, Elizabeth G. Campolongo, Matthew J. Thompson, Net Zhang, Jiaman Wu, Andrei Kopanev, and Alexander E. White
-- **Homepage:** https://imageomics.github.io/bioclip-2/
+- **Homepage:** [BioCLIP 2 Project Page](https://imageomics.github.io/bioclip-2/)
+- **Homepage:** [BioCLIP Ecosystem Site](https://imageomics.github.io/bioclip-ecosystem/)
 - **Repository:** [TreeOfLife-toolbox](https://github.com/Imageomics/TreeOfLife-toolbox)
 - **Paper:** [BioCLIP 2: Emergent Properties from Scaling Hierarchical Contrastive Learning](https://doi.org/10.48550/arXiv.2505.23883)
 
@@ -57,8 +62,14 @@ Image Classification, Zero-shot and few-shot Classification.
 /dataset/
     catalog.parquet
     embeddings/
-      txt_emb_species.json
-      txt_emb_species.npy
+        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/
@@ -73,7 +84,7 @@ Image Classification, Zero-shot and few-shot Classification.
             source=gbif/
 ```
 
-The `embeddings/` directory contains the [BioCLIP 2](https://huggingface.co/imageomics/bioclip-2) text embeddings for all 190,988,120 images with full taxonomic labels in the TreeOfLife-200M dataset. See [`embeddings/README.md`](embeddings/README.md) for more information.
+[BioCLIP 2](https://huggingface.co/imageomics/bioclip-2) and [BioCLIP 2.5 Huge](https://huggingface.co/imageomics/bioclip-2.5-vith14) 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`](http://github.com/Imageomics/pybioclip) functionality for versions 2.1.x and earlier. See [`embeddings/README.md`](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](https://github.com/Imageomics/TreeOfLife-toolbox/tree/main/docs#treeoflife200m-dataset-download-guide) 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.
@@ -93,9 +104,9 @@ Inside each shard is a collection of images (named `<uuid>.jpg`), for which each
 <uuid>.taxonomic_name.txt
 ```
 
-### Data Instances
+### Data Instances 
 
-Each image in this dataset is matched to the [7-rank Linnean taxonomy](https://www.britannica.com/science/taxonomy/The-Linnaean-system) and common name of the subject of the image (where available). Examples of these text types is provided below. 89% (190,988,120) of the images have full taxonomic labels (for more context, see discussion on labeling challenges in biodiversity data under [Considerations for Use](#bias-risks-and-limitations)). In addition to the biodiversity introduced (952,257 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](#data-curation-and-processing)) are provide in a table below.
+Each image in this dataset is matched to the [7-rank Linnean taxonomy](https://www.britannica.com/science/taxonomy/The-Linnaean-system) 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](#bias-risks-and-limitations)). 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](#data-curation-and-processing)) are provide in a table below.
 
 #### Text Types
 | Text Type | Example |
@@ -108,19 +119,19 @@ Each image in this dataset is matched to the [7-rank Linnean taxonomy](https://w
 ### Image Types (GBIF only)
 | Image Type | Number of Images |
 | :-------------------------- | :--------: |
-|Camera-trap                                     |617.8K        |
-|Citizen Science                                 |151M        |
+|Camera-trap                                     |595.6K        |
+|Citizen Science                                 |171.1M        |
 |Museum Specimen: Fungi                          |599.5K        |
-|Museum Specimen: Insect                         |7.9M          |
+|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: Uncategorized                  |1M            |
 |Museum Specimen: Vertebrate Zoology - Amphibians|36.2K         |
-|Museum Specimen: Vertebrate Zoology - Birds     |435.3K        |
+|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        |
+|Museum Specimen: Vertebrate Zoology - Others    |137K          |
 
 
 ### Data Fields
@@ -160,7 +171,7 @@ Each image in this dataset is matched to the [7-rank Linnean taxonomy](https://w
  - `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 (~2.5% of records). Non-GBIF records (EOL, BIOSCAN, FathomNet) have null values.
+ - `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](https://doi.org/10.48550/arXiv.2505.23883). 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](https://avibase.bsc-eoc.org/avibase.jsp).
  - `uuid`: unique identifier for the image in this dataset, links to other TreeOfLife-200M metadata.
@@ -218,13 +229,13 @@ This file is required along with `provenance.parquet` to fetch all the FathomNet
 
 ### Data Splits
 
-This entire dataset was used for training the model. 
+This entire dataset was used for training the model, [BioCLIP 2.5 Huge](https://huggingface.co/imageomics/bioclip-2.5-vith14) (similarly, all of [Revision a8f38b4](https://huggingface.co/datasets/imageomics/TreeOfLife-200M/tree/a8f38b4388579862c56ae57d6f094c2ac0e92e12) was used to train [BioCLIP 2](https://huggingface.co/imageomics/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](https://huggingface.co/imageomics/bioclip-2) was tested on the same 10 biologically-relevant benchmarks as [BioCLIP](https://huggingface.co/imageomics/bioclip#testing-data), though we used [NABirds](https://dl.allaboutbirds.org/nabirds) in place of [Birds 525](https://www.kaggle.com/datasets/gpiosenka/100-bird-species), since the latter is no longer available online. We also curated [IDLE-OO Camera Traps](https://huggingface.co/datasets/imageomics/IDLE-OO-Camera-Traps).
+[BioCLIP 2](https://huggingface.co/imageomics/bioclip-2) and [BioCLIP 2.5 Huge](https://huggingface.co/imageomics/bioclip-2.5-vith14) were tested on the same 10 biologically-relevant benchmarks as [BioCLIP](https://huggingface.co/imageomics/bioclip#testing-data), though we used [NABirds](https://dl.allaboutbirds.org/nabirds) in place of [Birds 525](https://www.kaggle.com/datasets/gpiosenka/100-bird-species), since the latter is no longer available online. We also curated [IDLE-OO Camera Traps](https://huggingface.co/datasets/imageomics/IDLE-OO-Camera-Traps).
 
  - [NABirds](https://dl.allaboutbirds.org/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](https://paperswithcode.com/dataset/meta-album): Specifically, we used the Plankton, Insects, Insects 2, PlantNet, Fungi, PlantVillage, and Medicinal Leaf datasets.
@@ -233,7 +244,7 @@ We used 11 biologically-relevant datasets for various species classification tes
 
 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](https://huggingface.co/imageomics/bioclip-2) model card and [our paper](https://doi.org/10.48550/arXiv.2505.23883).
+Other non-species classification tasks were also used for testing, and are described in the [BioCLIP 2](https://huggingface.co/imageomics/bioclip-2) and [BioCLIP 2.5 Huge](https://huggingface.co/imageomics/bioclip-2.5-vith14#testing-data) model card and [our paper](https://doi.org/10.48550/arXiv.2505.23883).
 
 ## Dataset Creation
 
@@ -241,13 +252,13 @@ Other non-species classification tasks were also used for testing, and are descr
 
 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](https://huggingface.co/datasets/imageomics/TreeOfLife-10M)) and the raw number of images available (i.e., from that of [BioTrove](https://huggingface.co/datasets/BGLab/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)](iucnredlist.org) as of [March 2025](https://nc.iucnredlist.org/redlist/content/attachment_files/2025-1_RL_Table_1a.pdf), with particularly strong representation of threatened species (77.1% of species across threatened categories are represented in TreeOfLife-200M). This coverage and the implications are discussed in greater detail in [our paper](https://doi.org/10.48550/arXiv.2505.23883).
+We have expanded coverage across the 2.14M described species estimated by [The International Union for Conservation of Nature (IUCN)](iucnredlist.org) as of [March 2025](https://nc.iucnredlist.org/redlist/content/attachment_files/2025-1_RL_Table_1a.pdf), 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](https://doi.org/10.48550/arXiv.2505.23883). 
 
 ### Source Data
 
 TreeOfLife-200M is composed of images from the following four core data providers:
 
-1. Global Biodiversity Information Facility ([GBIF](https://gbif.org)), which is a major biological data aggregator cataloging biodiversity data from citizen science sources (e.g., [iNaturalist](https://www.inaturalist.org/)), museum collections (e.g., from the [Smithsonian Institution](https://www.si.edu/), [Museum national d'Histoire naturelle](https://www.mnhn.fr/fr)), and camera trap collections (e.g., from the [Research Institute for Nature and Forest (INBO)](https://www.vlaanderen.be/inbo/en-gb/homepage/)). 
+1. Global Biodiversity Information Facility ([GBIF](https://gbif.org)), which is a major biological data aggregator cataloging biodiversity data from citizen science sources (e.g., [iNaturalist](https://www.inaturalist.org/) and [Observation.org](https://observation.org)), museum collections (e.g., from the [Smithsonian Institution](https://www.si.edu/), [Museum national d'Histoire naturelle](https://www.mnhn.fr/fr)), and camera trap collections (e.g., from the [Research Institute for Nature and Forest (INBO)](https://www.vlaanderen.be/inbo/en-gb/homepage/)). 
   - GBIF provides a [DOI](https://doi.org/10.15468/dl.bfv433) 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).
@@ -257,7 +268,7 @@ TreeOfLife-200M is composed of images from the following four core data provider
   - 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](https://opendata.eol.org/dataset/images-list/resource/f80f2949-ea76-4c2f-93db-05c101a2465c)). Taxonomy (`taxon.tab`) from "EOL Dynamic Hierarchy Active Version URL: https://editors.eol.org/uploaded_resources/00a/db4/dh21.zip" (at [dataset](https://opendata.eol.org/dataset/tram-807-808-809-810-dh-v1-1/resource/00adb47b-57ed-4f6b-8f66-83bfdb5120e8)).
   - Images used for the [Rare Species dataset](https://huggingface.co/datasets/imageomics/rare-species) were removed from this collection (by MD5) to avoid data leakage.
-  - Further filtereing and processing steps are described below.
+  - Further filtering and processing steps are described below.
 
 3. [BIOSCAN-5M](https://github.com/bioscan-ml/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](https://github.com/bioscan-ml/BIOSCAN-5M?tab=readme-ov-file#dataset-access) (specifically downloading through the [Google Drive link](https://drive.google.com/drive/u/1/folders/1Jc57eKkeiYrnUBc9WlIp-ZS_L1bVlT-0)). 
@@ -273,11 +284,11 @@ TreeOfLife-200M is composed of images from the following four core data provider
 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 |
+| Provider | Unique Taxa | Images | 
 | :---- | :---: |  :---: |
-| GBIF | 586,898 | 203.5M |
-| EOL | 48,748 | 5.3M |
-| BIOSCAN-5M | 2,900 | 5.2M |
+| GBIF | 569,225 | 222.6M |
+| EOL | 48,508 | 5.2M |
+| BIOSCAN-5M | 3,042 | 5.2M |
 | FathomNet | 251 | 37.9K |
 
 ### Data Curation and Processing
@@ -297,7 +308,7 @@ We noted the varied [image types](#image-types-gbif-only) included in the GBIF s
 
 Finally, images within ***citizen science occurrences*** are tested for similarity, and the mean pair-wise [BioCLIP](https://huggingface.co/imageomics/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](https://github.com/timesler/facenet-pytorch) on all images from GBIF and EOL to detect and remove images containing identifiable human faces. BIOSCAN-5M and FathomNet Database do not have images with human faces requiring filtering.
+We run [MTCNN](https://github.com/timesler/facenet-pytorch) 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**
 
@@ -309,21 +320,20 @@ The code for all of these processing steps, along with further details, is provi
 ### Annotations
 We standardized the taxonomic labels provided by the four core data providers to conform to a uniform [7-rank Linnean](https://www.britannica.com/science/taxonomy/The-Linnaean-system) structure. 
 
- #### Kingdom Counts
+#### Kingdom Counts 
 | Kingdom | Number of Images |
 | :-------------------------- | :--------: |
-|Animalia                            |102591559|
-|Plantae                             |101881537|
-|Fungi                               |7556801  |
-|Chromista                           |530234   |
-|Protozoa                            |184809   |
-|Bacteria                            |59109    |
-|Viruses                             |8587     |
-|Archaea                             |70       |
+|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](https://github.com/Imageomics/TaxonoPy) 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`](https://github.com/Imageomics/TaxonoPy) for more details on this process.
+Taxonomic labels (kingdom, phylum, etc.) were standardized across the various data sources using the [`TaxonoPy` package `v0.1.0`](https://github.com/Imageomics/TaxonoPy/releases/tag/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`](https://github.com/Imageomics/TaxonoPy) for more details on this process.
 
 #### Who are the annotators?
 
@@ -345,15 +355,18 @@ For instance, TreeOfLife-200M has a balanced representation between plants and a
 
 ![Graph displaying counts for most represented taxonomic classes, emphasizes imbalance](https://huggingface.co/datasets/imageomics/TreeOfLife-200M/resolve/main/visuals/taxa-class-imgs-most.png)
 
-Overall, 89% 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 30.78% 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. 
+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`](https://github.com/Imageomics/TaxonoPy)) resulted in less than 1% unresolved cases, e.g., `Metazoa` instead of `Animalia` or `incertae sedis` for kingdom:
+We note also that our taxonomic resolution ([`TaxonoPy`](https://github.com/Imageomics/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               |1060535  |
-|Archaeplastida               |33923    |
-|Metazoa                      |24858    |
-|NULL                         |5253     |
+| :----------------- | :--------------- |
+| incertae sedis  | 1056775          |
+| Archaeplastida     | 33891            |
+| Metazoa            | 24461            |
+| ...                |                  |
+| NULL               | 34               |
+| ...                |                  |
 
 
 ### Recommendations
@@ -386,26 +399,31 @@ If you use this dataset in your research, please cite both it and our associated
 **Data**
 ```
 @dataset{treeoflife_200m,
-  title = {{T}ree{O}f{L}ife-200{M} (Revision a8f38b4)}, 
+  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 = {2025},
+  year = {2026},
   url = {https://huggingface.co/datasets/imageomics/TreeOfLife-200M},
-  doi = {10.57967/hf/6786},
+  doi = {<update on release>},
   publisher = {Hugging Face}
 }
 ```
 
+Note that this version is updated from [Revision a8f38b4](https://huggingface.co/datasets/imageomics/TreeOfLife-200M/tree/a8f38b4388579862c56ae57d6f094c2ac0e92e12), 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](https://huggingface.co/imageomics/bioclip-2.5-vith14). Approximately 2.9 million images from [Revision a8f38b4](https://huggingface.co/datasets/imageomics/TreeOfLife-200M/tree/a8f38b4388579862c56ae57d6f094c2ac0e92e12) 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](https://huggingface.co/imageomics/bioclip-2.5-vith14) text embeddings were also added.
+
 Please also cite our paper:
 
 ```
-@article{gu2025bioclip,
-  title = {{B}io{CLIP} 2: Emergent Properties from Scaling Hierarchical Contrastive Learning}, 
-  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 = {2025},
-  eprint={2505.23883},
-  archivePrefix={arXiv},
-  primaryClass={cs.CV},
-  url={https://arxiv.org/abs/2505.23883},
+@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}
 }
 ```