Eval Recipes for upgrading to Gemini 2¶
This directory contains Eval Recipes for common GenAI tasks. The goal is to accelerate the process of upgrading to the latest version of Gemini and minimize the risk of regression. - An Eval Recipe is a minimalistic example of an automated evaluation that includes a prompt template and an evaluation dataset. - The included evaluation datasets are very small, which makes it possible to run each Eval Recipe in less than 1 minute. - Eval Recipes are lightweight and easy to learn - the entire configuration for most tasks fits on one screen. - Each Eval Recipe includes 3 alternative implementations: - Colab notebook based on Vertex GenAI Evaluation Service - Command line script based on Vertex GenAI Evaluation Service - Command line script based on Promptfoo - Eval Recipes can be customized by replacing the prompt template and the evaluation dataset.
Eval Recipe | Vertex AI Colab | Vertex AI Script | Promptfoo |
---|---|---|---|
Document QnA | view | view | view |
Summarization | view | view | view |
Text Classification | view | view | view |