Zero-Shot Image and Text Labeling in Label Genie
- Predictive AI
- H2O Cascade
This lesson covers zero-shot predictions in H2O Label Genie for two different data types. For an image dataset, it shows how a segment-anything-based model automatically draws segmentation masks around fish in each photo, and how keyboard shortcuts let a reviewer quickly approve or skip predictions to stay in a fast annotation flow. For a text dataset, it runs summarization on podcast transcript excerpts using a simpler, non-LLM default model, and points out where the summaries fall short, largely repeating opening sentences, setting up the next lesson on using large language models instead.
Key takeaways
- How zero-shot instance segmentation auto-draws masks around objects like fish without any prior training
- Why keyboard shortcuts for approving or skipping predictions help a reviewer move through a dataset quickly
- What limitation shows up when a simple non-LLM model is used to summarize long transcript text
- How exporting only the approved samples lets a first model be trained on a reviewed subset of data
Start owning your AI