LLM Response Ranking and Linguistic Refinement
Participated in a large-scale RLHF (Reinforcement Learning from Human Feedback) project aimed at improving chatbot conversational flow. The scope involved performing pairwise comparisons of model outputs to rank them based on helpfulness, honesty, and safety. Specific tasks included identifying hallucinations, correcting grammatical inconsistencies, and labeling intent across 2,500+ unique text strings. Adhered to strict quality measures including a 97% accuracy threshold on gold-standard audit tasks and consistent alignment with 50+ page project-specific guidelines.