optimize-anything

To optimize text artifacts without overfitting, you can use GEPA-based optimization loops with a dedicated validation set and transfer-aware evaluation strategies. Defining an evaluator contract and dataset alignment before spending budget helps ensure improvements are transfer-safe.

A seed_candidate is the initial text artifact input required to start iterative optimization in a GEPA workflow. It should be formatted as named, semantically meaningful sections like system_prompt or format_rules to preserve structure and enable targeted reflection.

You structure an evaluator contract by defining an ASI schema and deterministic scoring logic before running the optimizer. This ensures the evaluator provides actionable diagnostics and consistent scoring across the dataset during the iterative optimization loop.

Yes, you can run single-task, multi-task, and generalization workflows using a GEPAConfig composed from EngineConfig and ReflectionConfig. The configuration supports optional components like MergeConfig and RefinerConfig to handle complex multi-task optimization scenarios.

A validation set is optional but recommended for transfer-safe improvements during text artifact optimization. While you can run the optimizer with just a seed_candidate, dataset, and objective, providing a valset helps manage generalization checks and avoid overfitting.

Beyond default engine settings, GEPA optimization supports optional MergeConfig, RefinerConfig, and TrackingConfig components. You can also use public preset factories and seedless modes, with deeper API patterns available in the references directory for advanced configurations.