Update description to include the taxonomic updates done with TaxonoPy

README.md

@@ -4,6 +4,7 @@ language:
 - en
 - la
 pretty_name: TreeOfLife-10M
+description: Over 10 million images covering 454 thousand taxa in the tree of life used to train BioCLIP. This dataset of images of biological organisms paired with their associated taxonomic labels expands on the foundation established by existing high-quality datasets, such as iNat21 and BIOSCAN-1M, by further incorporating newly curated images from the Encyclopedia of Life (eol.org), which supplies most of TreeOfLife-10M’s data diversity.
 task_categories:
 - image-classification
 - zero-shot-classification
@@ -131,9 +132,9 @@ This dataset is a collection of images with associated text. The text matched to
   - `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.
-  - `common`: common name associated with the subject of the image where available. Otherwise, this is the scientific name (`genus-species`), else whatever subset of the taxonomic hierarchy is available (eg., `kingdom-phylum-class-order` or `kingdom-phylum-class-order-family`). All images have a non-null entry for this column.
+  - `common`: common name associated with the subject of the image where available from [GBIF](https://www.gbif.org/). Otherwise, this is the scientific name (`genus-species`), else whatever subset of the taxonomic hierarchy is available (eg., `kingdom-phylum-class-order` or `kingdom-phylum-class-order-family`). All images have a non-null entry for this column.
 
-Note that the `species` column occasionally has entries such as "sp. ___(get ex)" with some string following. This seems to be used to indicate the species is unknown, but various specimens/images are known to be the same species. Additionally, for `species` values containing an `x` between names, this is indicative of a hybrid that is a cross of the two species listed on either side of the `x`.
+Note that the `species` column occasionally has entries such as "sp. ___(get ex)" with some string following. This seems to be used to indicate the species is unknown, but various specimens/images are known to be the same species. Additionally, for `species` values containing an `x` between names, this is indicative of a hybrid that is a cross of the two species listed on either side of the `x`. Most, if not all, of these inconsistencies should have been resolved in the latest version. However, where a profile was not found to fit an entry and the taxonomic data was passed through from the source verbatim, such entries may occur. 
 
 ##### Text Types
 
@@ -147,7 +148,7 @@ Note that the `species` column occasionally has entries such as "sp. ___(get ex)
 `naming/`: these files provide the taxonomic name lookup for each data source in the reconstruction of the dataset.
 
 
-`species_level_taxonomy_chains.csv`: CSV with the ITIS taxonomic hierarchy, indicated as follows:
+`species_level_taxonomy_chains.csv`: CSV with the ITIS taxonomic hierarchy, indicated as follows. Note that the ITIS taxonomy was _not_ used for the latest version, only for the original, [Revision 69ca82e](https://huggingface.co/datasets/imageomics/TreeOfLife-10M/tree/69ca82e7481dcae0764f7c5452ea32b3a41cf620).
   - `hierarchy_string_tsn`: string of Taxonomic Serial Numbers (TSN)* for the names of the ranks provided from highest to lowest, connected by dashes (eg., `202422-846491-660046-846497-846508-846553-954935-5549-5550`).
   - `hierarchy_string_names`: string of the names of the ranks provided from highest to lowest, connected by arrows (eg., `Plantae->Biliphyta->Rhodophyta->Cyanidiophytina->Cyanidiophyceae->Cyanidiales->Cyanidiaceae->Cyanidium->Cyanidium caldarium`).
   - `terminal_tsn`: Taxonomic Serial Number (TSN)* of designated species (eg., `5550`).
@@ -211,10 +212,11 @@ Previously, the largest ML-ready biology image dataset was [iNat21](https://gith
 
 With this focus on species diversity and the need for high-quality images of biological organisms, we looked to the [Encyclopedia of Life Project (EOL)](https://eol.org/). EOL is an image aggregator that collaborates with a variety of institutions to source and label millions of images. After downloading 6.6M images from EOL, we were able to expand our dataset to cover an additional 440K taxa. 
 
-Insects (of the class Insecta with 1M+ species), birds (of the class Aves with 10K+ species) and reptiles (of the class Reptilia with 10K+ species) are examples of highly diverse subtrees with many more species than other taxonomic classes. This imbalance among subtrees in the tree of life present challenges in training a foundation model that can recognize extremely fine-grained visual representations of these classes. To help address this challenge for insects, we incorporated [BIOSCAN-1M](https://zenodo.org/doi/10.5281/zenodo.8030064), a recent dataset of 1M expert-labeled lab images of insects, covering 494 different families. The added variety of lab images, rather than in situ images (as in iNat21), further diversifies the _image_ distribution of TreeOfLife-10M.
+Insects (of the class Insecta with 1M+ species), birds (of the class Aves with 10K+ species) and reptiles (of the class Reptilia* with 10K+ species) are examples of highly diverse subtrees with many more species than other taxonomic classes. This imbalance among subtrees in the tree of life present challenges in training a foundation model that can recognize extremely fine-grained visual representations of these classes. To help address this challenge for insects, we incorporated [BIOSCAN-1M](https://zenodo.org/doi/10.5281/zenodo.8030064), a recent dataset of 1M expert-labeled lab images of insects, covering 494 different families. The added variety of lab images, rather than in situ images (as in iNat21), further diversifies the _image_ distribution of TreeOfLife-10M.
 
 Overall, this dataset contains approximately 454K unique taxonomic labels of the more than 2M recorded by [IUCN](iucnredlist.org) in 2022. To the best of our knowledge, this is still the most diverse and largest such ML-ready dataset available, hence our curation. 
 
+*<small>Note on Reptilia in the latest update: where a resolution was applied, Reptilia will _not_ show up. Where a profile was not found to fit an entry and the taxonomic data was passed through from the source verbatim, there are entries with class Reptilia. This affects approximately 1,207 records in the re-resolved 10M dataset.</small>
 
 ### Source Data
 
@@ -239,15 +241,10 @@ Species selected for the Rare Species dataset were removed from this dataset (se
 
 #### Annotation Process
 
-Annotations were primarily sourced from image source providers. 
+Annotations were initially sourced from the image providers, then their full 7-rank Linnean taxonomies were fed into [TaxonoPy](https://github.com/Imageomics/TaxonoPy) for alignment. 
 
-For iNat21 and BIOSCAN-1M images, the labels provided by those sources were used.
-- iNat21: iNaturalist English vernacular names and taxa were used.
-- BIOSCAN-1M: Linnaean taxonomic rankings were applied as labeled in the [BIOSCAN-1M dataset](https://zenodo.org/doi/10.5281/zenodo.8030064), which is all hand-labeled by experts. Note that the dataset provides other ranks (not considered in the 7-rank Linnaean taxonomy), such as tribe, which were not included in this dataset.
 
-For images from EOL, the scientific name (`genus-species`) was used to look up the higher-order taxa from the following sources as listed: BIOSCAN-1M metadata, EOL aggregate datasets (described below), then match this against the ITIS hierarchy for the higher-order taxa standardization. A small number of these are [homonyms](https://en.wikipedia.org/wiki/Homonym_(biology)), for which a list was generated to ensure proper matching of higher-order taxa (manual homonym resolution is in class `NameUpgrader` in the [naming script](https://github.com/Imageomics/bioclip/blob/main/src/imageomics/naming.py)). After these resources were exhausted, any remaining unresolved taxa were fed through the [Global Names Resolver (GNR) API](https://resolver.globalnames.org/api). Despite our efforts, we discovered after training that some hemihomonyms were mislabeled at higher-level taxa (family up to kingdom). This impacts approximately 0.1-0.2% of our data. We are in the process of developing a more robust solution to taxonomic labeling which will also account for re-naming (as is currently in process for many bird species). We intend to release a patch alongside the solution.
-
-This process allowed us to reach full taxa labels for 84% of images. To put this in perspective, 10% of images in TreeOfLife-10M are only labeled to the `family` level (no `genus-species` designations) as part of BIOSCAN-1M, so this places a cap on the taxa coverage. Taxonomic ranking also is not entirely standardized and agreed-upon throughout the biology community, so most gaps are more indicative of lack of consensus on label than missing information.
+This process allowed us to reach full taxa labels for 79.85% of images. To put this in perspective, 10% of images in TreeOfLife-10M are only labeled to the `family` level (no `genus-species` designations) as part of BIOSCAN-1M, so this places a cap on the taxa coverage. Taxonomic ranking also is not entirely standardized and agreed-upon throughout the biology community, so most gaps are more indicative of lack of consensus on label than missing information.
 
 
 #### Who are the annotators?
@@ -267,10 +264,11 @@ The hope is that this dataset could be helpful in conservation efforts or biodiv
 
 This dataset is imbalanced in its representation of various species with the greatest representation available for those in the phyla _Arthropoda_, _Tracheophyta_, and _Chordata_ (see our [interactive treemap from phylum to family](https://huggingface.co/imageomics/treeoflife-10m/raw/main/phyla_ToL_tree.html) for further details of this distribution). This class imbalance is both a result of availability of images and actual variance in class diversity. Additionally, as noted above, there are 2M+ estimated species according to [IUCN](iucnredlist.org), so overall taxonomic coverage is still limited (though it far surpasses the species diversity of other well-known animal datasets).
 
-Not all data is labeled to the species level, and some entries are more or less precise. For instance, the `species` column occasionally has entries such as "sp. ___(get ex)" with some string following. This seems to be used to indicate the species is unknown, but various specimens/images are known to be the same species. Additionally, for `species` values containing an `x` between names, this is indicative of a hybrid that is a cross of the two species listed on either side of the `x`. Due to the additional information provided about the higher order taxa, these labeling anomalies still present valuable information providing links between these classes.
+
+Not all data is labeled to the species level, and some entries are more or less precise. For instance, for `species` values containing an `x` between names, this is indicative of a hybrid that is a cross of the two species listed on either side of the `x`. Due to the additional information provided about the higher order taxa, these labeling anomalies still present valuable information providing links between these classes.
 
 
-As stated above, 84% of images have full taxa labels. However, due to the incomplete standardization and agreement on the taxonomic hierarchy throughout the biology community, most gaps are more indicative of lack of consensus on label than missing information. 
+As stated above, 79.85% of images have full taxa labels. However, due to the incomplete standardization and agreement on the taxonomic hierarchy throughout the biology community, most gaps are more indicative of lack of consensus on label than missing information. 
 
 Note that BIOSCAN-1M’s label granularity may still be limited for insects, as 98.6% of BIOSCAN-1M’s images are labeled to the family level but only 22.5% and 7.5% of the images have genus or species indicated, respectively. Lack of label granularity is an inherent challenge.
 
@@ -302,13 +300,16 @@ This dataset (the compilation) has been marked as dedicated to the public domain
 ```
 @dataset{treeoflife_10m,
   author = {Samuel Stevens and Jiaman Wu and Matthew J Thompson and Elizabeth G Campolongo and Chan Hee Song and David Edward Carlyn and Li Dong and Wasila M Dahdul and Charles Stewart and Tanya Berger-Wolf and Wei-Lun Chao and Yu Su},
-  title = {TreeOfLife-10M},
-  year = {2023},
+  title = {{TreeOfLife-10M (Revision )}},
+  year = {2025},
   url = {https://huggingface.co/datasets/imageomics/TreeOfLife-10M},
-  doi = {10.57967/hf/1972},
+  doi = {<update on new generation>},
   publisher = {Hugging Face}
 }
 ```
+
+Note that this version is updated from [Revision 69ca82e](https://huggingface.co/datasets/imageomics/TreeOfLife-10M/tree/69ca82e7481dcae0764f7c5452ea32b3a41cf620), which is the version used for the original BioCLIP model training ([BioCLIP Revision 7b4abf1](https://huggingface.co/imageomics/bioclip/tree/7b4abf1f6ee747c15de00c7d28a5e62990b5dabc)). This updated version resolves taxonomic alignment issues discovered in the first version. The taxonomic resolution was completed using [TaxonoPy](https://github.com/Imageomics/TaxonoPy), which was developed for [TreeOfLife-200M](https://huggingface.co/datasets/imageomics/TreeOfLife-200M).
+
 Please also cite our paper: 
 ```
 @inproceedings{stevens2024bioclip,
@@ -349,8 +350,22 @@ Please be sure to also cite the original data sources and all constituent parts
 For license and citation information by image, see our [license file](https://huggingface.co/datasets/imageomics/treeoflife-10m/blob/main/metadata/licenses.csv).
 
 
-- ITIS: Retrieved July, 20 2023, from the Integrated Taxonomic Information System (ITIS) on-line database, www.itis.gov, CC0
-https://doi.org/10.5066/F7KH0KBK
+
+Note that this version is updated from [Revision 69ca82e](https://huggingface.co/datasets/imageomics/TreeOfLife-10M/tree/69ca82e7481dcae0764f7c5452ea32b3a41cf620), which is the version used for the original BioCLIP model training ([BioCLIP Revision 7b4abf1](https://huggingface.co/imageomics/bioclip/tree/7b4abf1f6ee747c15de00c7d28a5e62990b5dabc)). [Revision 69ca82e](https://huggingface.co/datasets/imageomics/TreeOfLife-10M/tree/69ca82e7481dcae0764f7c5452ea32b3a41cf620) used ITIS for taxonomic resolution (Retrieved July, 20 2023, from the Integrated Taxonomic Information System (ITIS) on-line database, www.itis.gov, CC0
+https://doi.org/10.5066/F7KH0KBK). This update used [TaxonoPy](https://github.com/Imageomics/TaxonoPy):
+
+```bibtex
+@software{Thompson_TaxonoPy_2025,
+author = {Thompson, Matthew J. and Campolongo, Elizabeth G.},
+doi = {10.5281/zenodo.15499454},
+license = {MIT},
+month = may,
+title = {{TaxonoPy}},
+url = {https://github.com/Imageomics/TaxonoPy},
+version = {0.1.0-beta},
+year = {2025}
+}
+```
 
 
 ### Contributions