G-Eval
An evaluation metric method based on large models, in which an LLM performs fine-grained evaluation of outputs according to generated scoring steps. It originates from ideas in related academic papers.
Timeline (last 90 days)
When evaluation_steps is not passed, DeepEval's G-Eval metric automatically generates scoring steps based on criteria, causing the same response to potentially get opposite conclusions on re-runs, producing drift.
G-Eval originates from a paper's idea of having large models score step by step.
All Facts (2)
When evaluation_steps is not passed, DeepEval's G-Eval metric automatically generates scoring steps based on criteria, causing the same response to potentially get opposite conclusions on re-runs, producing drift.
50%UnverifiedG-Eval originates from a paper's idea of having large models score step by step.
50%