Google GAIA – Human Search Reasoning and Multimodal Dataset Creation
Internal Proprietary ToolingImageText GenerationRLHF
Worked under Turing as part of the Google GAIA (Gemini) data team to create high-quality datasets for training multimodal LLMs. Designed and annotated human reasoning sequences that replicate how users search, analyze, and synthesize information step by step across text, image, video, and audio inputs. Labeled reasoning chains, extracted relevant data, and wrote human-like solutions to teach models structured thinking. Validated model reasoning by comparing outputs with human-generated logic under RLHF and SFT frameworks. Ensured data accuracy, consistency, and diversity across modalities.
Worked under Turing as part of the Google GAIA (Gemini) data team to create high-quality datasets for training multimodal LLMs. Designed and annotated human reasoning sequences that replicate how users search, analyze, and synthesize information step by step across text, image, video, and audio inputs. Labeled reasoning chains, extracted relevant data, and wrote human-like solutions to teach models structured thinking. Validated model reasoning by comparing outputs with human-generated logic under RLHF and SFT frameworks. Ensured data accuracy, consistency, and diversity across modalities.
2024