Adala

Adala is a Python framework for AI engineers and data scientists who want to run data-labeling and evaluation workflows in their own process. It works with local pandas DataFrames and ground-truth examples, but the documented OpenAI and OpenRouter configurations send task inputs to cloud models. Those connections need internet access and an API key.

Examples guide the learning loop. Agents compare their work with known answers and use observations and feedback to refine task skills over repeated attempts. This is agent skill refinement, rather than training a neural network on your hardware. Developers can specify allowed labels and output constraints for tasks such as sentiment classification.

Adala fits Python notebooks and data-processing scripts. It applies learned skills to new records and returns predictions for further analysis. Skills also cover summarization, translation, text generation and answering questions from supplied context; a sequence of skills can handle work with several processing stages.

Separate teacher and student runtimes let developers use different models for learning and execution. OpenRouter provides a route to Claude and Gemini, while the LiteLLM runtime defaults to OpenAI and accepts other configured providers or custom OpenAI-compatible endpoints. A locally managed dataset can still produce a remote model request for each record.

The framework is open source under Apache 2.0. Developers can extend its skills, runtimes and agent types to fit their own labeling tasks and feedback environments.