Adala

What is Adala?

Adala is a framework for building autonomous agents specialized in data processing and data-labeling tasks. Its name expands to Autonomous Data Labeling Agent. An Adala agent can acquire one or more labeling skills through an iterative learning process shaped by its environment, observations, and reflections. The user supplies that environment through a ground-truth dataset.

The framework separates several concepts involved in agent-based labeling. Agents provide the main interface, environments supply data and ground-truth signals, skills define labeling behavior, memories store and retrieve information, and runtimes connect those skills to an LLM provider. The documentation uses a sentiment-classification task to demonstrate how an agent learns from labeled examples and then predicts labels for new text.

Adala is distributed as a Python package and can be used as a standalone library in a Python notebook. The documented quickstart uses an OpenAI chat runtime, a static pandas-based environment, and a classification skill.

How to Use Adala

1. Install the Python package with pip install adala.
2. Set the OPENAI_API_KEY environment variable for the runtime used in the documented quickstart.
3. Create a ground-truth dataset and connect it to a StaticEnvironment.
4. Define a skill, such as ClassificationSkill, with instructions, labels, and input and output templates.
5. Configure the available runtimes and select the default runtime for the agent.
6. Run iterative learning with a chosen accuracy threshold, then apply the trained skill to test data with the agent's run method.

Core Features

• Autonomous skill learning: Agents can acquire one or more data-labeling skills through repeated learning iterations.
• Ground-truth environments: Users provide datasets that supply the environment and reference signal for learning.
• Configurable skills: Labeling behavior is defined through skill instructions, labels, and input and output templates.
• LLM runtimes: Skills execute through configured runtimes, with the documentation showing an OpenAI chat runtime.
• Agent memory: The framework includes memories for storing and retrieving information used by agents.
• Python workflow: Adala can run as a standalone library in notebooks and process pandas DataFrames.

Use Cases

• Text classification: Data teams define categories and train an agent from labeled examples before applying it to new text.
• Sentiment labeling: Practitioners use a classification skill and a ground-truth dataset to assign positive, negative, or neutral sentiment labels.
• Custom data-labeling agents: Developers combine environments, skills, memories, and runtimes to implement labeling workflows for their own datasets.

Frequently Asked Questions

What does Adala stand for?

Adala stands for Autonomous Data Labeling Agent.

What supplies the learning signal for an Adala agent?

Users define an environment by providing a ground-truth dataset, which influences learning together with the agent's observations and reflections.

How does Adala execute a labeling skill?

Each skill runs through a configured runtime. The quickstart demonstrates OpenAIChatRuntime as the default runtime.

Can Adala be used directly in a Python notebook?

Yes. The documentation's quickstart uses Adala as a standalone library in a Python notebook with pandas DataFrames.