What is evalship?
Evalship provides engineering teams with a robust framework for evaluating LLM-driven software by performing thorough audits of evaluation suites and scrutinizing pull requests. It carefully investigates scorers and judges for possible inaccuracies, identifies gaps in coverage, and points out evaluations that could lead to erroneous results, all while delivering insights complete with code references and suggestions for enhancements. The initial setup streamlines a coding agent's task of configuring GitHub Actions and creating a setup pull request. For future pull requests, Evalship assesses the results against the default branch, detecting regressions, flaky evaluations, and any changes to prompts or model calls that lack sufficient evaluation, while also providing a status check for merge gating. It accommodates results from a variety of supported test runners and LLM evaluation tools. Evaluations occur within the customer's GitHub Actions framework using their designated model API keys, with Evalship concentrating on reviewing the results rather than executing the evaluations. Furthermore, teams have the ability to review their evaluation history and receive weekly performance reports to monitor advancements and enhancements over time. This thorough methodology guarantees that maintaining software quality is prioritized at every stage of the development process, ultimately fostering a culture of continuous improvement.