--- dataset_info: - config_name: full features: - name: session dtype: string - name: filename dtype: string - name: sampling_type dtype: string - name: image dtype: image - name: site dtype: int32 - name: day dtype: int32 - name: plant dtype: string - name: level dtype: string - name: view dtype: string - name: healthy dtype: string - name: rust dtype: string - name: leaf_miner dtype: string - name: other_insect dtype: string - name: mechanical_damage dtype: string - name: other_remarks dtype: string - name: expert_healthy dtype: string - name: expert_rust dtype: string - name: expert_leaf_miner dtype: string - name: expert_other_insect dtype: string - name: expert_mechanical_damage dtype: string - name: expert_confidence dtype: string - name: expert_other_remarks dtype: string - name: expert_notes dtype: string splits: - name: dorsal num_bytes: 11501554442 num_examples: 821 - name: ventral num_bytes: 11474680169 num_examples: 821 - name: both num_bytes: 22640962726.17 num_examples: 1642 download_size: 44513598362 dataset_size: 45617197337.17 - config_name: opportunistic features: - name: session dtype: string - name: filename dtype: string - name: sampling_type dtype: string - name: image dtype: image - name: site dtype: int32 - name: day dtype: int32 - name: plant dtype: string - name: level dtype: string - name: view dtype: string - name: healthy dtype: string - name: rust dtype: string - name: leaf_miner dtype: string - name: other_insect dtype: string - name: mechanical_damage dtype: string - name: other_remarks dtype: string - name: expert_healthy dtype: string - name: expert_rust dtype: string - name: expert_leaf_miner dtype: string - name: expert_other_insect dtype: string - name: expert_mechanical_damage dtype: string - name: expert_confidence dtype: string - name: expert_other_remarks dtype: string - name: expert_notes dtype: string splits: - name: dorsal num_bytes: 5943664030 num_examples: 423 - name: ventral num_bytes: 5930372510 num_examples: 423 - name: both num_bytes: 11874036540 num_examples: 846 download_size: 23749429714 dataset_size: 23748073080 - config_name: systematic features: - name: session dtype: string - name: filename dtype: string - name: sampling_type dtype: string - name: image dtype: image - name: site dtype: int32 - name: day dtype: int32 - name: plant dtype: string - name: level dtype: string - name: view dtype: string - name: healthy dtype: string - name: rust dtype: string - name: leaf_miner dtype: string - name: other_insect dtype: string - name: mechanical_damage dtype: string - name: other_remarks dtype: string - name: expert_healthy dtype: string - name: expert_rust dtype: string - name: expert_leaf_miner dtype: string - name: expert_other_insect dtype: string - name: expert_mechanical_damage dtype: string - name: expert_confidence dtype: string - name: expert_other_remarks dtype: string - name: expert_notes dtype: string splits: - name: dorsal num_bytes: 5557890412 num_examples: 398 - name: ventral num_bytes: 5544307659 num_examples: 398 - name: both num_bytes: 11102198071 num_examples: 796 download_size: 20796610703 dataset_size: 22204396142 configs: - config_name: full data_files: - split: dorsal path: full/dorsal-* - split: ventral path: full/ventral-* - split: both path: full/both-* default: true - config_name: opportunistic data_files: - split: dorsal path: opportunistic/dorsal-* - split: ventral path: opportunistic/ventral-* - split: both path: opportunistic/both-* - config_name: systematic data_files: - split: dorsal path: systematic/dorsal-* - split: ventral path: systematic/ventral-* - split: both path: systematic/both-* license: cc-by-nc-4.0 task_categories: - image-classification - image-segmentation tags: - plants - images - segmentation - classification - "clidemia hirta" - "leaf damage" - "biocontrol agents" - biology size_categories: 1K This dataset is aimed at the image multi-classification and segmentation of various leaf damage types caused by biocontrol agents. The dataset contains images of both the dorsal and ventral side of *Clidemia Hirta* leaves, that were all collected in January 2025 near Hilo (Hawaii), in dirt trails along *Steinback Highway*. *Clidemia Hirta* is a highly invasive plant on the island of Hawaii (Big Island). ## Dataset Configurations and Splits Three configurations of the dataset can be loaded from the HuggingFace `datasets` library. Each configuration can be further subdivided into three analytical splits. ### The Systematic Configuration The systematic configuration will download to your local machine the subset of the dataset corresponding to the leaves that were collected by following our systematic sampling protocol (see the paper for more details). The `dorsal` split will only download the images corresponding to the dorsal side of these leaves, the `ventral` split will only download the images corresponding to the ventral side, and the `both` split will include both sides of the images. Here's how to load each of the three available splits : ```python from datasets import load_dataset # Dorsal Split dorsal_dataset = load_dataset("imageomics/invasive_plants_hawaii", name="systematic", split="dorsal") # Ventral Split ventral_dataset = load_dataset("imageomics/invasive_plants_hawaii", name="systematic", split="ventral") # Combined Ventral and Dorsal Split combined_dataset = load_dataset("imageomics/invasive_plants_hawaii", name="systematic", split="both") ``` ### The Opportunistic Configuration The opportunistic configuration will download to your local machine the subset of the dataset corresponding to the leaves that were collected by following our opportunistic sampling protocol (see the paper for more details). The `dorsal` split will only download the images corresponding to the dorsal side of these leaves, the `ventral` split will only download the images corresponding to the ventral side, and the `both` split will include both sides of the images. Here's how to load each of the three available splits : ```python from datasets import load_dataset # Dorsal Split dorsal_dataset = load_dataset("imageomics/invasive_plants_hawaii", name="opportunistic", split="dorsal") # Ventral Split ventral_dataset = load_dataset("imageomics/invasive_plants_hawaii", name="opportunistic", split="ventral") # Combined Ventral and Dorsal Split combined_dataset = load_dataset("imageomics/invasive_plants_hawaii", name="opportunistic", split="both") ``` ### The Full Configuration The full configuration will download the full dataset (i.e., both opportunistically and systematically sampled leaves) to your local machine . The `dorsal` split will only download the images corresponding to the dorsal side of the leaves, the `ventral` split will only download the images corresponding to the ventral side, and the `both` split will include both sides of the images. Here's how to load each of the three available splits : ```python from datasets import load_dataset # Dorsal Split dorsal_dataset = load_dataset("imageomics/invasive_plants_hawaii", name="full", split="dorsal") # Ventral Split ventral_dataset = load_dataset("imageomics/invasive_plants_hawaii", name="full", split="ventral") # Combined Ventral and Dorsal Split combined_dataset = load_dataset("imageomics/invasive_plants_hawaii", name="full", split="both") ``` ## Dataset Details ### Dataset Description - **Curated by:** David Carlyn, Catherine Villeneuve, Leonardo Viotti, Ellyn Bitume, Chuck Stewart, Leanna House - **Language:** English - **Repository:** [Invasives Code](https://github.com/Imageomics/invasives-project) ### Baseline Damage Classification Benchmark **F1-Score** | Model / Damage | Healthy | Leaf Miner | Fungus | Other Insect | Mechanical | | :--------------: | :-----: | :--------: | :----: | :----------: | :--------: | | **ResNet-50** | 69.3% | 66.8% | 49.7% | 68.9% | 60.7% | | **ResNext** | 73.5% | 66.4% | 62.0% | 66.2% | 59.9% | | **ConvNext** | **87.2%** | **78.0%** | **76.2%** | 73.2% | **67.8%** | | **ViT-base** | 48.5% | 69.5% | 47.4% | 71.2% | 66.0% | | **MaxViT** | 81.5% | 70.3% | 72.9% | **74.9%** | 52.1% | | **Conv ViT** | 70.7% | 71.9% | 66.5% | 60.7% | 65.1% | **ROC-AUC** | Model / Damage | Healthy | Leaf Miner | Fungus | Other Insect | Mechanical | | :--------------: | :-----: | :--------: | :----: | :----------: | :--------: | | **ResNet-50** | 84.0% | 82.4% | 68.9% | 73.6% | 70.4% | | **ResNext** | 88.1% | 82.3% | 71.1% | 77.0% | 70.6% | | **ConvNext** | **97.1%** | **88.2%** | **86.1%** | **87.1%** | **73.8%** | | **ViT-base** | 81.2% | 84.6% | 65.9% | 81.4% | 70.8% | | **MaxViT** | 93.6% | 83.8% | 82.1% | 86.4% | 71.3% | | **Conv ViT** | 85.7% | 81.1% | 69.8% | 71.9% | 73.2% | ### Imaging Sessions The leaves were imaged during 12 distinct imaging sessions (marked as `session` in the dataset). The session `session_1_16_2025` was our initial test imaging setup and we don't recommend using the images associated with this session to train any leaf damage classification or segmentation model. We used only the images from session 2 to session 12 to train our model. These images were carefully taken in a lab-controlled setting, with an intense bright light, on a 1/4 inch white grid paper. The four red dots can be used to remove distortion in the images, and the color card can be used to calibrate the colors. Two different phones were used to image the leaves, namely : * **Sessions 1, 2, 3, 4, 5, 6, 9, 11 :** iPhone 13 * **Sessions 7, 8, 10, 12 :** Pixel 3 ### Leaf ID Each leaf can be uniquely identified by the following informations : * **Imaging session :** The name of the imaging session during which the dorsal and ventral pictures of the leaf were taken, marked in the `session` column of the dataset. * **Site ID :** The ID of the site (marked in the `site` column of the dataset) where the leaf was collected. * **Plant ID :** The ID of the plant (marked in the `plant` column of the dataset) from which the leaf was collected. For opportunistically sampled leaves, the plant ID will be noted as `Exp`. * **Level :** The level of the plant (marked in the `level` column of the dataset) from which the leaf was collected. For leaves sampled opportunistically, the value of the `level` column is equal to `None`. For the leaves sampled systematically, the value of `level` column can be either * `L` : Leaf was sampled from the bottom level of the plant * `M` : Leaf was sampled from the middle of the plant * `H` : Leaf was sampled from the top of the plant * **Index :** An integer number to help identify each leaf individually, marked in the `index` column of the dataset ### Extracting Images and Distinguishing Between Dorsal and Ventral Sides The images corresponding to the dorsal and ventral side of each can be easily accessed by using the following informations: * **Filename :** The filename of the image, marked in the `filename` column of the dataset * **View :** The side of the leaf (`D` for dorsal, `V` for ventral) associated to the `filename` image. This information is marked in the `view` column of the dataset ### Damage Types (Classification Labels) A given leaf can be completely healthy (which will be indicated in the `healthy` column of the dataset), or show a combination of the following types of damage: * **Fungus Damage :** Damage caused by *Colletotrichum gloeosporioides*, an antrachnose fungus that was released in Hawaii as a biocontrol agent for the first time in 1997 ([Conant et al., 2013](https://www.fs.usda.gov/rm/pubs_other/rmrs_2013_conant_p001.pdf))([Killgore, 2002](http://www.hear.org/pcsu/techreports/pdfs/pcsutechreport00129killgore.pdf))([Conant, 2002](http://www.hear.org/pcsu/techreports/pdfs/pcsutechreport00129conant.pdf)). Infected leaves will have dark brown lesions that resemble rust spots. The presence or absence of fungus damage for a given leaf will be indicated in the `rust` column of the dataset. * **Mechanical Damage :** Mechanical damage caused notably by human/animal manipulation, pesticides, or cars. Affected leaves will have dark irregular lesions on them. These lesions have a different shape and are darker than those caused by *C. gloeosporioides*. The presence or absence of mechanical damage for a given leaf will be indicated in the `mechanical_damage` column of the dataset. * **Leaf Miner Damage :** Damage caused by leaf miner insects. The larvae of *Lius Poseidon*, released in Hawaii in 1988 as a biocontrol agent for *C. Hirta* ([Conant et al., 2013](https://www.fs.usda.gov/rm/pubs_other/rmrs_2013_conant_p001.pdf)) ([Conant, 2002](http://www.hear.org/pcsu/techreports/pdfs/pcsutechreport00129conant.pdf)), is a notable example. The larvae creates a distinctive brown serpentine shape, that widens as the larvae grows. The presence or absence of leaf miner damage for a given leaf will be indicated in the `leaf_miner` column of the dataset. * **Other Insects Damage :** Damage caused by other insects, notably leaf defoliation. The presence or absence of other insect damage for a given leaf will be indicated in the `other_insect` column of the dataset. For a given leaf, each damage type column (`healthy`, `rust`, `other_insect`, `leaf_miner`, `mechanical_damage`) will be marked by either Yes, Maybe, or No. The column `other_remarks` can also include additional notes about the leaf damage(s) that were observed by the human who labelled it. ### Expert Validation Expert validation on the classification labels of the leaves is available for a subset of the dataset. For a given leaf, the damage types observed by the expert are marked in the columns `expert_healthy`, `expert_rust`, `expert_leaf_miner`, `expert_mechanical_damage` and `expert_other_insect`. The expert's confidence level in their annotations is stored in the `expert_confidence` column. Additional remarks from the expert on the observable damage of the leaf can be found in the `expert_other_remarks` and `expert_notes` columns. ## Supported Tasks and Leaderboards Our dataset supports image multi-class classification, as well as image segmentation. See our paper for current benchmarks. ## Dataset Structure ``` /invasive_plants_hawaii/ docs/ field_imaging_and_extraction_protocol.pdf lab_imaging_protocol.pdf metadata/ full_dataset.csv full_opportunistic_dataset.csv full_systematic_dataset.csv full_dataset_dorsal.csv systematic_dataset_dorsal.csv opportunistic_dataset_dorsal.csv full_dataset_ventral.csv systematic_dataset_ventral.csv opportunistic_dataset_ventral.csv protocol/ imaging_sessions_metadata.csv sites_metadata_gps.csv trips_metadata.csv full/ both-00000-of-00046.parquet dorsal-00000-of-00024.parquet ventral-00000-of-00023.parquet ... both-00045-of-00046.parquet dorsal-00023-of-00024.parquet ventral-00022-of-00023.parquet systematic/ both-00000-of-00023.parquet dorsal-00000-of-00012.parquet ventral-00000-of-00012.parquet ... both-00023-of-00024.parquet dorsal-00011-of-00012.parquet ventral-00011-of-00012.parquet opportunistic/ both-00000-of-00024.parquet dorsal-00000-of-00012.parquet ventral-00000-of-00012.parquet ... both-00023-of-00024.parquet dorsal-00011-of-00012.parquet ventral-00011-of-00012.parquet images/ session_1_1_16_2025/ {img_name}.png session_2_1_18_2025/ {img_name}.png ... session_12_1_28_2025/ {img_name}.png ``` The folder `docs` contains the .pdf files detailing our leaf imaging protocols. The folder `metadata` contains the .csv files liking each image to their labels, for each configuration (`full`, `opportunistic`, `systematic`) and for each split (`dorsal`, `ventral`, `both`). The subfolder `protocol`, located inside the `metadata` folder, contains the .csv files summarizing the metadata associated with each imaging session, each sampling site and each sampling trip. The folders `full`, `opportunistic` and `systematic` contains the parquet files that can be loaded when using the `load_dataset` function from the HuggingFace datasets library. Each parquet file is approximately between 400 and 800 MB. The `image` folder contains the images of the leaves, splitted between 12 imaging sessions. The archived images used to produce our results for leaf reconstruction are available in `images/Dataset_Leaf_Reconstruction.zip` while images used to produce our RCNN results are in `images/rcnn_images.zip`. ## Data Instances This dataset is a collection of *Clidemia Hirta* leaves images paired with associated damage classes, identified either by 1) a computer scientist or 2) a computer scientist and an expert. There are **821 leaves**, each imaged on both their dorsal and ventral side (**1642** images in total, 821 dorsal images, 821 ventral images). Of these 821 leaves, 423 of them were opportunistically sampled, and the remaining 398 were collected according to our systematic sampling protocol. ## Data Fields The files `full_dataset.csv`, `full_opportunistic_dataset.csv.csv`, `full_systematic_dataset.csv`, `full_dataset_dorsal.csv`, `systematic_dataset_dorsal.csv`, `opportunistic_dataset_dorsal.csv`, `full_dataset_ventral.csv`, `systematic_dataset_ventral.csv` and `opportunistic_dataset_ventral.csv` all share the same following metadata with each image in the dataset : - `session` : The session during which the leaf was imaged - `filename` : The filename of the .PNG image associated with the leaf in the session folder. Unique ID. - `site` : The site ID where the leaf was collected - `day` : The day of when the leaf was collected (a day of January 2025) - `plant` : The ID of the plant from which the leaf was collected - `level` : The level of the plant from which the leaf was collected - `index` : The ID number associated with the leaf - `view` : The side from which the leaf was taken (dorsal or ventral) - `healthy` : `Yes` if the leaf is surely healthy, `Maybe` if the leaf is potentially healthy, `No` if the leaf is surely not healthy - `rust` : `Yes` if the leaf has fungus damage, `Maybe` if the leaf has potentially fungus damage, `No` if the leaf doesn't show any sign of fungus damage - `leaf_miner` : `Yes` if the leaf has been damaged by a leaf miner, `Maybe` if the leaf was potentially damaged by a leaf miner, `No` if the leaf wasn't damaged by a leaf miner - `other_insect` : `Yes` if the leaf was damaged by an insect that isn't a leaf miner, `Maybe` if the leaf was potentially damaged by an insect that is not a leaf miner, `No` if the leaf was not damaged by an insect that is not a leaf miner - `mechanical_damage` : `Yes` if the leaf has mechanical damaged, `Maybe` if there is potentially some signs of mechanical damage, `No` if there is no sign of mechanical damage on the leaf - `other_remarks` : Other remarks about the leaf written by the human who labelled it - `expert_healthy` : If the leaf was classified or not as healthy by the expert. Can be `Yes`, `Maybe`, `No` or `None` (if the leaf was not labelled by the expert) - `expert_rust` : If the expert observed any sign of fungus damage on the leaf. Can be `Yes`, `Maybe`, `No` or `None` (if the leaf was not labelled by the expert) - `expert_other_insect` : If the expert observed any sign of other insect damage on the leaf. Can be `Yes`, `Maybe`, `No` or `None` (if the leaf was not labelled by the expert) - `expert_leaf_miner` : If the expert observed any sign of leaf miner damage on the leaf. Can be `Yes`, `Maybe`, `No` or `None` (if the leaf was not labelled by the expert) - `expert_confidence` : How confident the expert was in their labels - `expert_other_remarks` : Additional remarks from the expert on the damage they observed on the leaf - `expert_notes` : Additional remarks from the expert on the damage we can observe on the leaf - `sampling_type` : Describes the sampling strategy that was used to collect the leaf. Can be `opportunistic` or `systematic` The `metadata/protocol/trips_metadata.csv` file contains information about each field data collection trip during which the leaves were collected : - `trip_id` : The ID of the trip. ID created based on order of trip taken (e.g., ID 1 means it was the first trip taken) - `site_id` : The sampling site ID we visited during that trip - `date` : The date we did the field collection trip - `start_time` : The time when we started the field collection trip - `end_time` : The time when we ended the field collection trip - `notes` : Relevant notes about the sampling trip The `metadata/protocol/imaging_sessions_metadata.csv` file contains information about each leaf imaging session : - `id` : The ID of the imaging session (between 1 and 12). ID created based loosely on order of imaging (e.g., ID 1 means it was the first imaging session) - `location` : The location where we did the imaging session - `camera` : The camera used to take pictures of the leaves - `photographer` : The individual who took the pictures of the leaves during that imaging session - `distance_of_camera_to_specimen`: the estimated distance from the camera to the leaf specimen when the photograph was taken. - `session_start_time` : Start time of the imaging session - `session_end_time` : End time of the imaging session - `notes` : Relevant notes about the imaging session The `metadata/protocol/sites_metadata_gps.csv` file contains information about site visisted for collection : - `site_name` : The nickname of the visted site. - `id` : ID created based on order of site visited (e.g., ID 1 means it was the first site visited) - `latitude` : The latitude to the entrance of the site visited. - `longitude` : The longitude to the entrance of the site visited. - `elevation_estimates` : Elevation of the site visited. Only contains partial information (some elevation data missing). - `notes` : Relevant notes about the site to help with locating. ## Dataset Creation This dataset was compiled as part of the field component of the Experiential Introduction to AI and Ecology Course run by the Imageomics Institute and the AI and Biodiversity Change (ABC) Global Center. This field work was done on the island of Hawai'i January 15-30, 2025. ## Curation Rationale This dataset was created with the goal of developping a deep learning-based tool that would allow conservationists in Hawaii to automate their monitoring of the impact of biocontrols on invasive plants species distribution in Hawaii. We selected the species *Clidemia Hirta* as our field collection target because its abundance, but our protocols can be deployed and our algorithms can be trained on any similar kind of datasets. ## Source Data The raw images were downloaded from an iPhone 13 and a Pixel 3, then transformed into a .PNG format. The images taken with an iPhone 13 were converted from their native .HEIF format to .PNG using the [pillow-heif](https://pypi.org/project/pillow-heif/) library. Please feel free to reach out to us if you are interested in getting the raw images. ## Data Collection and Processing Please see `docs/field_imaging_and_extraction_protocol.pdf` and `docs/lab_imaging_protocol.pdf` for more information on our data collection, processing and imaging protocols. ## Annotations ### Annotation process Classification annotations were done after imaging, and were done after carefully looking at the leaf. We recorded our labels into a shared Google Drive Sheet, that we later cleaned and transformed into a standardized .csv format. ### Who are the annotators? * **Expert Annotator :** Ellyn Bitume * **Student Annotators :** David Carlyn, Catherine Villeneuve and Leonardo Viotti ### Personal and Sensitive Information *Clidemia Hirta* is an highly invasive plant of the island of Hawaii. If you are interesting in using the GPS locations of our field sites to reproduce our protocol/collect more leaves, please follow the appropriate car/clothes/shoe cleaning guidelines in order to prevent the further spread of *Clidemia Hirta* on the island. ## Considerations for Using the Data Please use the citation bellow if you end up using our dataset. ### Bias, Risks, and Limitations Our systematic sampling protocol is probably not 100% bias free. Please be careful if you use this data to infer any relationship with e.g. damage and environmental covariates, and please reach out to us if you would like more detailed informations about the limitations of our systematic sampling protocol. ### Recommendations We don't recommend using our systematic sampling protocol as is to infer relationship between *Clidemia Hirta* biocontrols and environmental covariates. Careful statistical processing is required. ## Licensing Information This project is licensed under Creative Commons Attribution-NonCommercial 4.0 International [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/deed.en). ## Citation **Data:** ``` @dataset{invasive_plants_hawaii_dataset, author = {David Carlyn* and Leonardo Viotti* and Catherine Villeneuve* and Kazi Sajeed Mehrab and Ellyn Bitume and Leanna House and Chuck Stewart}, title = {Hawaii Leaf Damage Dataset}, year = {2025}, url = {https://huggingface.co/datasets/imageomics/invasive_plants_hawaii}, publisher = {Hugging Face} } ``` * is used to indicate equal contribution. ## Acknowledgements This work was supported by both the [Imageomics Institute](https://imageomics.org) and the [AI and Biodiversity Change (ABC) Global Center](https://abcresearchcenter.org/). The Imageomics Institute is funded by the US National Science Foundation's Harnessing the Data Revolution (HDR) program under [Award #2118240](https://www.nsf.gov/awardsearch/showAward?AWD_ID=2118240) (Imageomics: A New Frontier of Biological Information Powered by Knowledge-Guided Machine Learning). The ABC Global Center is funded by the US National Science Foundation under [Award No. 2330423](https://www.nsf.gov/awardsearch/showAward?AWD_ID=2330423&HistoricalAwards=false) and Natural Sciences and Engineering Research Council of Canada under [Award No. 585136](https://www.nserc-crsng.gc.ca/ase-oro/Details-Detailles_eng.asp?id=782440). This dataset draws on research funded by the Social Sciences and Humanities Research Council. 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 or Natural Sciences and Engineering Research Council of Canada. We acknowledge the support of the [U.S. Forest Service](https://www.fs.usda.gov/), more specifically the researchers associated with the [Pacific Southwest Research Station](https://research.fs.usda.gov/psw) (Institute of Pacific Islands Forestry), for their support and collaboration in this research. Data collection and protocols were notably done in close collaboration with Dr. Ellyn Bitume, a research entomologist for the U.S. Forest Service. We also acknowledge the support of the [National Ecological Observatory Network](https://www.neonscience.org/) (NEON), a program sponsored by the U.S. National Science Foundation (NSF) and operated under cooperative agreement by Battelle. We especially thank the team associated with the [Pu'u Maka'ala Natural Area Reserve](https://www.neonscience.org/field-sites/puum), for helping us with logistics and equipments. ## Dataset Card Authors Catherine Villeneuve and David Carlyn ## Dataset Card Contact Catherine Villeneuve can be reached out through firstname.lastname@mila.quebec David Carlyn can be reached out through lastname.1@osu.edu