Download an annotated dataset
Overview
At any point in an annotation task, you can download the already approved annotated samples. You do not need to fully annotate an imported dataset to download already annotated samples (approved samples).
An array of datasets labeled in H2O Label Genie are supported in H2O Hydrogen Torch and H2O LLM Studio.
Instructions
To download an annotated dataset (approved samples), consider the following instructions:
- On the H2O Label Genie navigation menu, click Annotation tasks.
- In the annotation tasks table, double-click the row where the annotation task you want to download is located.
- Click the Export tab.
- In the Export approved samples list, select Download ZIP.
Note
H2O Label Genie downloads a zip file containing the annotated dataset in a particular format depending on the dataset's annotation task. To learn more, see Downloaded dataset formats.
- Zero-shot learning models: H2O Label Genie generates zero-shot predictions for specific annotation tasks that can be downloaded or exported to H2O Drive.
- To learn how to download a dataset's zero-shot predictions, see Download a dataset's zero-shot predictions
- To learn how to export a dataset's zero-shot predictions to H2O Drive, see Export a dataset's zero-shot predictions to H2O Drive
- Export an annotated dataset to H2O Drive: You can export an annotated dataset (approved samples) to H2O Drive. To learn more, see Export an annotated dataset to H2O Drive.

Downloaded dataset formats
Each zip file is named after the task, its data type, and its problem type, for example reviews_text_classification.zip for a text classification task named reviews. The problem type matches the task type, except for entity recognition (token_classification), summarization (sequence_to_sequence), and generative AI (llm). The zip file contains a label file named labels.csv, labels.pq, or labels.json and, for image and audio tasks, an images/ or audios/ folder with the approved files.
Text classification
A downloaded text classification dataset (with approved samples) follows the following dataset format: A zip file (1) containing a CSV file (2):
folder_name.zip (1)
│ └───csv_name.csv (2)
- A zip file.
- A CSV file containing a text column and either a label column or one
label_CLASS_NAMEcolumn for each class, whereCLASS_NAMEis the class name:- text: The text column contains the text input
- label: In a single-label task with two or more classes, the label column contains the labels attributed to the texts specified in the text column
label_CLASS_NAME: In a multi-label task, or a binary task (which has only one class), eachlabel_CLASS_NAMEcolumn contains1if the text has that class and0if it doesn't
H2O Hydrogen Torch supports the downloaded dataset of a text classification annotation task. In particular, H2O Hydrogen Torch supports the downloaded dataset for a text classification experiment.
Text regression
A downloaded text regression dataset (with approved samples) follows the following dataset format: A zip file (1) containing a CSV file (2):
folder_name.zip (1)
│ └───csv_name.csv (2)
- A zip file.
- A CSV file containing a text and label column:
- text: The text column contains the text input
- label: The label column contains the labels attributed to the texts specified in the text column
H2O Hydrogen Torch supports the downloaded dataset of a text regression annotation task. In particular, H2O Hydrogen Torch supports the downloaded dataset for a text regression experiment.
Text-entity recognition
A downloaded text entity recognition dataset (with approved samples) follows the following dataset format: A zip file (1) containing a .pq file (2):
folder_name.zip (1)
│ └───pq_name.pq (2)
- A zip file.
- A
.pqfile containing a raw_text, text, and label column:- raw_text: The raw_text column contains the original text, which H2O Label Genie uses to match labels to samples when you import the file into an annotation task
- text: The text column contains the text split into tokens at spaces and at the start and end of each entity
- label: The label column contains one label for each token in the text column: the token's entity, or
Ofor a token outside any entity
Text summarization
A downloaded text summarization dataset (with approved samples) follows the following dataset format: A zip file (1) containing a CSV file (2):
folder_name.zip (1)
│ └───csv_name.csv (2)
- A zip file.
- A CSV file containing a text and label column:
- text: The text column contains the text input
- label: The label column contains the summaries attributed to the texts specified in the text column
Text-generative AI
A downloaded text-generative AI dataset (with approved samples) follows the following dataset format: A ZIP file (1) containing a CSV file (2):
folder_name.zip (1)
│ └───csv_name.csv (2)
- A zip file.
- A CSV file containing the following columns:
- llm_input: The final prompts sent to the large language model (LLM) as an input query
- llm_output: The answers provided to the query (llm_input)
- original columns: All the original columns
H2O LLM Studio supports the downloaded dataset of a text-generative AI annotation task.
Image classification
A downloaded image classification dataset (with approved samples) follows the following dataset format: A zip file (1) containing a CSV file (2) and an image folder (3).
folder_name.zip (1)
│ └───csv_name.csv (2)
│ │
│ └───image_folder_name (3)
│ └───name_of_image.image_extension
│ └───name_of_image.image_extension
│ └───name_of_image.image_extension
│ ...
- A zip file.
- A CSV file containing an image_path column and either a label column or one
label_CLASS_NAMEcolumn for each class, whereCLASS_NAMEis the class name:- image_path: The image_path column specifies the location of the image
- label: In a single-label task with two or more classes, the label column contains the labels attributed to the images specified in the image_path column
label_CLASS_NAME: In a multi-label task, or a binary task (which has only one class), eachlabel_CLASS_NAMEcolumn contains1if the image has that class and0if it doesn't
- Image folder that contains all the images specified in the image_path column.
H2O Hydrogen Torch supports the downloaded dataset of an image classification annotation task. In particular, H2O Hydrogen Torch supports the downloaded dataset for an image classification experiment.
Image regression
A downloaded image regression dataset (with approved samples) follows the following dataset format: A zip file (1) containing a CSV file (2) and an image folder (3).
folder_name.zip (1)
│ └───csv_name.csv (2)
│ │
│ └───image_folder_name (3)
│ └───name_of_image.image_extension
│ └───name_of_image.image_extension
│ └───name_of_image.image_extension
│ ...
- A zip file.
- A CSV file containing an image_path and label column:
- image_path: The image_path column specifies the location of the image
- label: The label column contains the labels attributed to the images specified in the image_path column
- Image folder that contains all the images specified in the image_path column
H2O Hydrogen Torch supports the downloaded dataset of an image regression annotation task. In particular, H2O Hydrogen Torch supports the downloaded dataset for an image regression experiment.
Object detection
A downloaded object detection dataset (with approved samples) follows the following dataset format: A zip file (1) containing a .pq file (2) and an image folder (3).
folder_name.zip (1)
│ └───pq_name.pq (2)
│ │
│ └───image_folder_name (3)
│ └───name_of_image.image_extension
│ └───name_of_image.image_extension
│ └───name_of_image.image_extension
│ ...
- A zip file.
- A
.pqfile with one row for each approved image, containing an image_path, class_id, x_min, x_max, y_min, and y_max column:- image_path: The image_path column specifies the location of the image
- class_id: The class_id column contains the list of class names of the image's bounding boxes
- x_min, x_max, y_min, and y_max: The x_min, x_max, y_min, and y_max columns contain the lists of bounding box coordinates in pixels, in the order of the class_id column
- Image folder that contains all the images specified in the image_path column.
H2O Hydrogen Torch supports the downloaded dataset of an object detection task. In particular, H2O Hydrogen Torch supports the downloaded dataset for an image object detection experiment.
Image instance segmentation
A downloaded image instance segmentation dataset (with approved samples) follows the following dataset format: A zip file (1) containing a JSON file (2) and an image folder (3).
folder_name.zip (1)
│ └───json_name.json (2)
│ │
│ └───image_folder_name (3)
│ └───name_of_image.image_extension
│ └───name_of_image.image_extension
│ └───name_of_image.image_extension
│ ...
- A zip file.
- A JSON file containing the labels in COCO format, with the following keys:
- images: The ID, path (file_name), width, and height of each approved image
- annotations: One entry for each polygon, with its points in pixels (segmentation), the image ID (image_id), and the class ID (category_id)
- categories: The ID and name of each class in the rubric
- Image folder that contains all the images specified in the JSON file.
H2O Hydrogen Torch supports the downloaded dataset of an image instance segmentation annotation task. In particular, H2O Hydrogen Torch supports the downloaded dataset for an image instance segmentation experiment.
Audio classification
A downloaded audio classification dataset (with approved samples) follows the following dataset format: A zip file (1) containing a CSV file (2) and an audio folder (3).
folder_name.zip (1)
│ └───csv_name.csv (2)
│ │
│ └───audio_folder_name (3)
│ └───name_of_audio.audio_extension
│ └───name_of_audio.audio_extension
│ └───name_of_audio.audio_extension
│ ...
- A zip file.
- A CSV file containing an audio_path column and either a label column or one
label_CLASS_NAMEcolumn for each class, whereCLASS_NAMEis the class name:- audio_path: The audio_path column specifies the location of the audio
- label: In a single-label task with two or more classes, the label column contains the labels attributed to the audios specified in the audio_path column
label_CLASS_NAME: In a multi-label task, or a binary task (which has only one class), eachlabel_CLASS_NAMEcolumn contains1if the audio has that class and0if it doesn't
- Audio folder that contains all the audios specified in the audio_path column.
H2O Hydrogen Torch supports the downloaded dataset of an audio classification annotation task. In particular, H2O Hydrogen Torch supports the downloaded dataset for an audio classification experiment.
Audio regression
A downloaded audio regression dataset (with approved samples) follows the following dataset format: A zip file (1) containing a CSV file (2) and an audio folder (3).
folder_name.zip (1)
│ └───csv_name.csv (2)
│ │
│ └───audio_folder_name (3)
│ └───name_of_audio.audio_extension
│ └───name_of_audio.audio_extension
│ └───name_of_audio.audio_extension
│ ...
- A zip file.
- A CSV file containing an audio_path and label column:
- audio_path: The audio_path column specifies the location of the audio
- label: The label column contains the labels attributed to the audios specified in the audio_path column
- Audio folder that contains all the audios specified in the audio_path column.
H2O Hydrogen Torch supports the downloaded dataset of an audio regression annotation task. In particular, H2O Hydrogen Torch supports the downloaded dataset for an audio regression experiment.
- Submit and view feedback for this page
- Send feedback about H2O Label Genie to cloud-feedback@h2o.ai