May 2023 Summaries

Predibase has announced the general availability and free trial of its low-code AI platform, designed to simplify the development of AI-powered applications by enabling developers and data scientists to build and fine-tune models without needing to master complex ML frameworks. The platform employs a "Declarative ML" approach, inspired by technologies like dbt and Terraform, allowing users to define desired outcomes while the platform automates the creation of end-to-end ML pipelines. This approach, already adopted by companies like Uber and Meta, allows for rapid iteration and deployment of models, drastically reducing development time. Predibase also introduces the Data Science Copilot, a tool that aids iterative model improvement, and supports privately hosted, customized Large Language Models (LLMs) that can be tailored to specific enterprise workflows. The platform offers both a managed cloud and a virtual private cloud option for deployment, catering to the needs of various organizations, from startups to Fortune 500 companies.

A tutorial guides users through building a Named Entity Recognition (NER) model for molecular biology text using Predibase, a low-code declarative machine learning platform. NER, an essential NLP task, identifies and categorizes entities like proteins or genes within text, crucial for organizing molecular biology data. The tutorial uses the BioNLP/JNLPBA dataset to train the model, starting with data preparation and preprocessing into a format suitable for Predibase. Users create a model repository and configure the model using Predibase's interface, leveraging the Ludwig framework to simplify the process. Initial training on 10% of the dataset achieves 87% accuracy, prompting iterative improvements by adjusting parameters like sample ratio and encoder settings, ultimately reaching 92% accuracy. The tutorial demonstrates three methods to operationalize the model, including deploying via Predibase's managed endpoint, exporting the model, or using the Predictive Query Language (PQL) for batch inference. Predibase significantly reduces the complexity and time required to develop NER models, making advanced tasks more accessible.