Is it possible to add a corrected response (e.g. from a SME) and not just a Good/bad binary feedback to the annotation? The idea is to curate a preference dataset out of this and possibly use for fine-tuning or adding to examples in the prompt (adapative in-context learning).
