Training data for reward models

Reward models are trained on examples with human judgments of quality. Useful records include QA scores and reviewer notes, approved versus rejected drafts, and customer satisfaction ratings. They provide real preference signals from expert reviewers, applied with consistent rubrics over time.

Records for reward models have personal details removed, and the supplier approves every release. A rights review runs on QA scores, reviewer notes, drafts and ratings before a sample is prepared. The prepared sample is shared under NDA only after supplier approval.

Reward models are scoped through a request describing the records, volume and uses you need. SourceX matches suitable suppliers in the U.S., Canada, the EU, the UK, Australia, New Zealand, India, Singapore, South Africa, Latin America, the UAE and Saudi Arabia and runs a rights review. The license sets allowed uses, term and deletion, and resale and re-identification are barred.