Reward Hacking Detection Preference Data Reviewer

AuraOne, Inc.; United States; Remote; USD 34,000 - 55,000; Full time

Job summary

AuraOne, Inc. is seeking a Reward Hacking Detection Preference Data Reviewer to remotely review prompts and responses against AuraOne's quality rubric. You will compare paired outputs, label edge cases, and provide structured feedback the modeling team can use to retrain.

As part of AI data review, you will label hallucinations and unsafe content, capture ambiguous prompts, and calibrate against gold-standard examples weekly. Reliable async availability (10+ hours) is required.

Qualifications

• Prior evaluation, annotation, or human-rater experience on reward hacking detection preference data evaluation or adjacent content.
• Comfort applying multi-page rubrics consistently across long batches.
• Clear written reasoning that names the issue and the rubric clause being applied.
• Strong attention to detail and the ability to flag when a prompt itself is the problem.
• Reliable async availability for at least 10 hours per week.

Responsibilities

• Evaluate reward hacking detection preference data evaluation model outputs against a versioned rubric and assign severity tags for Reward Hacking Detection Preference Data Reviewer assignments.
• Compare paired responses and pick the stronger answer with a written rationale.
• Label hallucinations, instruction-following failures, and unsafe content with structured tags.
• Capture ambiguous prompts and route them back to the program team for rubric updates.
• Maintain reviewer-quality scores by calibrating against gold-standard examples each week.
• Document recurring failure modes so the modeling team can target them in the next training run.

Skills

Model output evaluation

Rubric-based annotation

Severity tagging

Inter-rater calibration

Reward Hacking Detection Preference  "

Preference ranking

RLHF

Rater calibration

Reward

Hacking

Detection

Job description

Reward Hacking Detection Preference Data Reviewer is a remote evaluation track for reviewing reward hacking detection preference data evaluation prompts and responses against AuraOne's quality rubric. Reviewers compare paired outputs, label edge cases, and write the kind of structured feedback the modeling team can use to retrain.

Why this role matters

AI data reviewers help turn reward hacking detection preference data evaluation outputs into auditable labels, rationales, and regression cases for AuraOne Human Data.

Responsibilities

• Evaluate reward hacking detection preference data evaluation model outputs against a versioned rubric and assign severity tags for Reward Hacking Detection Preference Data Reviewer assignments.
• Compare paired responses and pick the stronger answer with a written rationale.
• Label hallucinations, instruction-following failures, and unsafe content with structured tags.
• Capture ambiguous prompts and route them back to the program team for rubric updates.
• Maintain reviewer-quality scores by calibrating against gold-standard examples each week.
• Document recurring failure modes so the modeling team can target them in the next training run.

Qualifications

• Prior evaluation, annotation, or human-rater experience on reward hacking detection preference data evaluation or adjacent content for Reward Hacking Detection Preference Data Reviewer work.
• Comfort applying multi-page rubrics consistently across long batches.
• Clear written reasoning that names the issue and the rubric clause being applied.
• Strong attention to detail and the ability to flag when a prompt itself is the problem.
• Reliable async availability for at least 10 hours per week.

Example tasks

• Compare two reward hacking detection preference data evaluation model responses to the same prompt and pick the stronger one with rationale.
• Tag an unsafe response with the correct policy category and severity.
• Audit a 50-row batch for rubric consistency and report drift to the program lead.
• Propose a rubric clarification after spotting a recurring failure mode.

Nice to have

• Background in linguistics, content moderation, or trust & safety review.
• Experience with inter-rater agreement metrics and calibration cycles.
• Domain expertise that lets you spot subject-matter errors automated checks miss.

Skills

• Model output evaluation
• Rubric-based annotation
• Severity tagging
• Inter-rater calibration
• Reward Hacking Detection Preference Data evaluation
• Preference ranking
• RLHF
• Rater calibration
• Reward
• Hacking
• Detection

Work model

Remote — US-eligible. Remote · Independent specialist contractor. Employment type: CONTRACTOR. Applicants must be authorized to work from US.

Compensation

Hourly rate confirmed after the interview process.