offline data warehouse

Large-scale sentiment labeling

Run business-specific labels across warehouse-scale text without routing every row through a frontier model.

Result

A post-trained 30B open model labeled 39,962 comments with frontier-like aggregate rates at 4.4x lower cost than Sonnet and 50x lower cost than Opus.

The public benchmark reports aggregate label agreement, sparse intent labels, sample size, cost, and non-claims so the evidence is inspectable.

Metrics
39,962 comments labeled
4.4x lower cost than Sonnet on the measured table
50x lower cost than Opus on the measured table
frontier-like aggregate rates on dense sentiment labels
Route

Workflow

The task is high-volume text labeling where the business needs consistent labels across many rows, not a bespoke answer for each row.

Quality contract

The label set, sparse intent fields, disagreement rows, and frontier control slice define whether a cheaper route is close enough to promote.

Optimization path

Understudy uses the frontier model where it creates new information: hard cases, adjudication, rubric repair, and control sampling. The specialist handles the repeatable layer.

Buyer fit

This is a fit when a team has warehouse-scale text, domain-specific labels, and enough volume for a specialist route to repay the optimization work.

Proof