HumanSignal launches open-source Adala framework for autonomous data annotation agents

HumanSignal, the company behind the widely used open-source data annotation tool Label Studio, has now launched the Adala open-source framework for autonomous data labeling agents. Previously known as Heartex, HumanSignal underwent a name change in June 2023 to draw attention to its core value proposition of involving humans in the loop of machine learning (ML) training. Data annotation is a fundamental activity in training models and has historically been a labor-intensive process. With Label Studio, data scientists have access to tools for annotating different types of data, including text and video. With the rapid development of machine learning, HumanSignal aims to shape the future of reliable and efficient data processing through the new open-source framework, Adala. Adala stands for Autonomous Data Labeling Agent and utilizes AI agents in a novel way to accelerate and improve the data annotation process. Michael Malyuk, co-founder and CEO of HumanSignal, said, "We started thinking about what it means to build a trustworthy AI agent. Adala is our response, aimed at helping build autonomous and reliable agents specifically for data processing tasks." How Adala Accelerates the Data Annotation Process The Adala agent aims to learn and improve data tasks such as classification and annotation by providing a ground truth dataset. The ground truth dataset serves as the foundation for defining data labels and can be developed using Label Studio technology. Malyuk explained that in the Adala framework, there is a concept of an environment that essentially defines how the agent learns through interaction with the ground truth dataset. The Adala agent interacts with the environment, learns from it, and becomes a prediction engine after multiple learning iterations. In the initial use case of Adala, these predictions are applied to the rest of the unlabeled dataset. Adala agent is driven by what Malyuk calls a "runtime" component, which is essentially a large language model (LLM). The runtime executes the tasks specified for the agent and provides responses. Nikolai Liubimov, Chief Technology Officer of HumanSignal, explained that the architecture of the Adala framework requires some form of storage, typically a vector database. He noted that the process of obtaining data labels applicable to new data is similar in many ways to retrieval augmented generation (RAG) in vector databases. Adala Goes Beyond Data Annotation Malyuk pointed out that the Label Studio user community has been requesting various automation features. While the initial functionality of Adala is data annotation, he emphasized that it can serve as a general-purpose agent for various data processing tasks. By releasing the Adala project as open-source, he hopes that users can contribute ideas and code for extending Adala. Malyuk said, "In a year, there will be agents of different types with different skills, capable of interacting and receiving feedback from different types of environments. It's a very powerful approach, and we want to share it with a wider community."