Skip to main content
Version: v0.4.0

Release notes

v0.4.0 | October 25, 2023​

Overview​

Because nobody likes to label data, H2O Label Genie has integrated with h2oGPT and OpenAI large language models (LLMs) to expedite the labeling process.

The new version introduces several new and improved features to enhance the user experience while streamlining the labeling process. At the core of the new version, the following three major points:

  • New annotation task: H2O Label Genie now supports a text-generative AI annotation task. This new annotation task with h2oGPT and OpenAI LLMs makes it possible to create more complex annotation tasks, for example, around translation, question answering, and summarization.
  • LLMs: h2oGPT and OpenAI LLMs are now available for text summarization and text-generative AI annotation tasks.
  • Box to polygon: By default now, for an image instance segmentation task, H2O Label Genie offers a box to polygon feature that enables you to improve the efficiency of manual labeling. The new feature enables you to encapsulate an object class with a rectangular that is immediately replaced with a more precise polygon that captures all the borders of the object class inside the rectangular.

Annotation tasks​

  • New: H2O Label Genie now supports a text-generative AI annotation task.
    • Why?: This new update further expands the mission to offer a platform to generate labeled datasets supported in H2O Hydrogen Torch, H2O LLM Studio, and other products in and out of the H2O AI Cloud. This new annotation task with h2oGPT and OpenAI LLMs support makes it possible to create more complex annotation tasks, for example, around translation, question answering, and summarization. To learn more, see Supported annotation tasks: Text-generative AI.
  • New: By default now, for an image instance segmentation annotation task, H2O Label Genie offers a box to polygon feature that enables you to improve the efficiency of manual labeling. A box to polygon feature enables you to encapsulate an object class with a rectangular that is immediately replaced with a more precise polygon that captures all the borders of the object class inside the rectangular.
    • Why?: This new update expedites the annotation process for image instance segmentation tasks by precisely annotating segmentation masks while eliminating the tedious task. In other words, the new update helps to add masks quickly when a zero-shot model misses something. To learn more, see Segment objects.
  • New: Text summarization and text-generative AI annotation tasks now support the following large language models(LLMs): h2oGPT, OpenAI GPT-3.5/4.
  • Improvement: H2O Label Genie has improved the ability to annotate the same dataset with others. Now, you can change the default setting to enable others to work on the same annotation task.
    • Why?: This new update enables you to collaborate with others in real time to speed up the annotation process further. To learn more, see Collaboration settings
  • Improvement: H2O Label Genie now prioritizes zero-shot predictions for in-view annotation task samples.
    • Why?: Calculating zero-shot predictions requires a lot of computational resources; therefore, zero-shot predictions are queued. If the priority of a particular annotation task changes to high, you can rest assured H2O Label Genie will respond with a similar urgency.
  • Improvement: Now, H2O Label Genie selects a new color for a particular new label for annotation tasks utilizing colors to differentiate labels. You no longer need to select a color for a new label manually.
    • Why?: This new update expedites the process of creating an annotation task rubric.
  • Improvement: H2O Label Genie now automatically assigns a label to an in-view sample based on generated zero-shot predictions for annotation tasks supporting zero-shot learning models.
    • Why?: This new update expedites the labeling and approval of a particular sample by tapping into usually accurate zero-shot predictions.
  • Bugfix: The following reported issue has been fixed: It was reported that zero-shot predictions (probabilities) for multi-label text classification annotation tasks were being normalized to 100% by H2O Label Genie.
  • Bugfix: The following reported issue has been fixed: It was reported that the following workflow was breaking the application:
    1. Begin annotating a dataset.
    2. Simultaneously, generate zero-shot predictions.
    3. Once the first batch of zero-shot predictions is ready, review and return to an annotated sample for further evaluation.

Datasets​

  • New: H2O Label Genie imputes missing values in a text dataset by filling in missing values with the following: No text.
    • Why?: In hopes of consistency, H2O Label Genie keeps all provided rows during import despite having missing values.

UI & UX​

  • New: The new H2O Label Genie v0.4.0 comes with a new logo.
  • New: When importing a dataset through the S3 connector, H2O Label Genie now renders a progress bar displaying an estimated time for H2O Label Genie to download and validate the dataset.
    • Why?: Importing large datasets through the S3 connector can take a while.
  • Improvement: Now, on the Annotation tasks page, you can view the progress of an annotation task producing zero-shot predictions.
    • Why?: This update brings about the ability to quickly view the status of all zero-shot predictions for all annotation tasks supporting it.
  • Improvement: Now, in the Dashboard tab of an annotation task, you can view a dataset sample's thumbnail image and labels.
    • Why?: This new update facilitates the process of overviewing and reviewing ongoing annotation tasks.

Documentation​

Feedback