LLM Safety & Instruction Alignment Annotator – Multimodal (Text, Image, Video Transcript)
Took the lead in conducting a 3-month, 500+ prompt response annotation project of text, image and video transcript annotation. Transcribed 5 clips of video, performed binary safety classification, 5-point Likert ratings, bounding box annotation (25 images), and pairwise ranking. Followed a 12-page rubric (based on HH/Moderation standards) and reached 95% consistency in the self-audits and recorded 45 edge cases with justifications. All quality measures – blind re-runs, inter-rater checks, precision coordinates – represent production level AI training pipelines.