Snorkel developer advocate Kobie Crawford presents a collaboration with UC Berkeley on training a four-billion-parameter model for financial question answering. The team uses expert-created data and reinforcement learningReinforcement learning trains an AI system to choose actions using feedback about the results of its earlier choices. in a tool-use environment.
Crawford contrasts a larger model that guesses table names with the smaller trained model discovering tables, inspecting schemas and correcting errorsModel reliability is the degree to which an AI model produces dependable behavior across repeated, varied and operationally realistic use.. He reports task-specific benchmark gainsEvaluation is the systematic assessment of a system's behaviour or results against explicit criteria., including transfer from simpler training questions to harder multi-table questions.
Watch on YouTube
