Data Labeling Batches

Scale AI turns raw data into labeled training sets through its data engine. A Beam agent readies a labeling batch by reading the incoming dataset, applying the task spec and quality rules your ML lead approved, and submitting it to the right labeling queue. It records batch status and writes results back once labels return. Batches with unclear instructions, sensitive content, or an edge case the spec does not cover are not pushed through; the agent holds them and routes the questions to a person, who clarifies the guidelines before the batch goes out for labeling.

Annotation Quality Review

Returned annotations in Scale AI carry quality and confidence signals. A Beam agent reads each completed batch, compares its scores against the acceptance bar your team set, and marks the set as accepted or sent back for rework according to that rule. It updates the project record and notifies the owner. Sets that land near the threshold, show inconsistent labeling, or touch a class your policy treats as sensitive are not auto-accepted; the agent flags them and hands them to a reviewer, who inspects the samples and makes the final accept-or-reject decision by hand.

Generative Model Data Prep

The Scale generative platform assembles enterprise data for model tuning and evaluation. A Beam agent gathers the approved source records, applies the filtering and formatting rules your team defined, and stages the dataset for a tuning or evaluation run. It logs what was included and updates the run record. Records that carry personal data, unclear licensing, or content the policy restricts are never staged automatically; the agent separates them and routes the question to a data owner, who reviews the source and decides whether it can be used for training.