Coreset selection
Coreset selection chooses a small, representative subset of a dataset that approximates the full set for training or labeling. It cuts labeling and compute costs while retaining most of the information, by keeping diverse, informative samples and dropping redundant ones.
In this article
Building visual or physical AI?
Let's talk.
What is coreset selection?
Coreset selection picks a compact subset of a dataset that behaves, for the purposes of training or evaluation, much like the whole thing. Rather than sampling at random, it favors samples that are diverse and informative while dropping redundant ones, so a fraction of the data captures most of its value.
It is a practical answer to the fact that labeling and training on everything is often unnecessary and expensive.
Key takeaways
- A coreset is a small subset that approximates the full dataset.
- It keeps diverse, informative samples and drops redundant ones.
- It reduces labeling and compute cost with minimal performance loss.
How it works
Selection methods use embeddings to measure similarity and coverage, then choose points that span the dataset well, for example by maximizing diversity or coverage of the embedding space. Some approaches also weight by informativeness, such as model uncertainty. The result is a labeling or training budget spent where it counts.
Before you pick a training subset, see what the full dataset covers
Coreset selection picks a small subset that trains a model nearly as well as the full dataset. Many methods work on embeddings: keep samples that cover the space and drop the ones crowded together. The quiet problems:
- a dense cluster of near-identical images that random sampling keeps picking from
- a rare class that a coverage-based method drops because it sits near a common one
- class balance that holds in the full set and breaks in the subset
In Voxel51, Open Images 10K has 10,000 images balanced across 20 classes, with embeddings, and you can ask the Voxel51 Agent to explore the embeddings or show the class distribution.
Open Images 10K
10,000 images
Open Images V7 images balanced across 20 everyday classes, with boxes, instance masks, image labels, points, and relationships.
Try this: Ask the Voxel51 Agent to explore embeddings, then find the densest cluster and decide how many of its images you would keep.
Explore the images in Voxel51
Free with a work or .edu email.
Why it matters
When labeling is the bottleneck, choosing what to label matters as much as how well you label it. Coreset selection directs limited annotation budget toward the most useful samples, and it connects closely to active learning, where model feedback further sharpens the choice.
Frequently asked questions
How is coreset selection different from random sampling?
Random sampling ignores content, while coreset selection deliberately chooses diverse, informative samples that better represent the dataset.
How does it relate to active learning?
Both aim to label the most useful data. Active learning selects iteratively using model feedback, while coreset selection can be a one-shot, coverage-driven choice.
Related terms

Director of Growth
Jesse Mostipak is the Director of Growth at Voxel51, where the work is helping humans find and trust what the brand knows, and teaching the Google knowledge graph and the LLMs answering on their behalf to do the same. That question, how knowledge gets built inside a system, is one Jesse has been chasing for years. Earlier versions of it ran through a New York City high school science classroom, data science and machine learning, and developer relations at Kaggle, Posit (formerly RStudio), and Baseten. The answer doesn't change much depending on whether the learner is a teenager, a software engineer, or a knowledge graph. Jesse holds a Master's in Education from CUNY Hunter College. LinkedIn
See all articles by Jesse Mostipak
Last updated November 11, 2026