HumanSignal, previously known as Heartex, has unveiled the Adala open source framework, aimed at revolutionizing the data labeling process for machine learning (ML) training. By integrating humans into the loop, HumanSignal aims to improve the reliability and efficiency of data processing in the rapidly evolving AI landscape.
The Importance of Data Labeling
Data labeling is a crucial step in training ML models and has traditionally been a labor-intensive task. With the introduction of Label Studio, data scientists gained access to versatile tools for labeling various data types, including text and video.
Introducing Adala: The Autonomous Data Labeling Agent
Adala, short for Autonomous Data Labeling Agent, is a groundbreaking approach that leverages AI agents to accelerate and enhance the data labeling process. Michael Malyuk, cofounder and CEO of HumanSignal, explained the motivation behind Adala, stating, “We wanted to create a reliable AI agent that can be trusted. Adala is our response, designed to build autonomous agents dedicated to data processing tasks.”
Adala agents are trained to handle classification and labeling tasks by utilizing ground truth datasets. The foundation of a ground truth dataset is established using the Label Studio technology. Within the Adala framework, an agent interacts with the environment, learns from it, and evolves into a prediction engine after multiple iterations of learning.
Nikolai Liubimov, CTO of HumanSignal, highlighted the role of a runtime in powering Adala agents. The runtime, based on a large language model (LLM), executes the designated tasks and provides responses. Additionally, the Adala framework architecture necessitates the use of storage, typically in the form of a vector database. Retrieving a data label for new data follows a similar process to Retrieval Augmented Generation (RAG) in LLMs.
“The initial capability enabled by Adala is data labeling, but it can be a generalized agent for a variety of data processing tasks,” emphasized Michael Malyuk. As the project is open source, HumanSignal encourages users to contribute ideas and code tailored to their specific requirements. Malyuk envisions a future where different types of agents with diverse skills can interact and receive feedback from various environments.
HumanSignal aims to share this powerful approach with the wider community and anticipates significant advancements in Adala’s capabilities in the coming year.
