#rfms — Public Fediverse posts

"The culmination of this research and foresight led my team and me to create the first Relational Foundation Model (RFM) for business data. Its purpose is to enable machines to reason directly over structured data, to understand how entities, such as customers, transactions, and products, connect. By knowing the relationships between these entities, we then enable users to make accurate predictions from those specific relationships and patterns.

Unlike LLMs, RFMs have been designed for structured relational data. RFMs are pretrained on a number of (synthetic) datasets as well as on a number of tasks over structured business data. Like LLMs, RFMs can be simply prompted to produce instant responses to a wide variety of predictive tasks over a given database, all without task-specific or database-specific training.

We wanted a system that could learn directly from how real databases are structured, and without all the usual manual setup. To make that possible, we treated each database like a graph: tables became node types, rows turned into nodes, and foreign keys linked everything together. This way, the model could actually “see” how things like customers, transactions, and products connect and change over time.

At the heart of it, the model combines a column encoder with a relational graph transformer. Every cell in a table is turned into a small numerical embedding based on what kind of data it holds, whether it’s a number, category, or a timestamp. The Transformer then looks across the graph to pull context from related tables, which helps the model adapt to new database schemas and data types.

For users to input which predictions they’d like to make, we built a simple interface called Predictive Query Language (PQL). It lets users describe what they want to predict, and the model takes care of the rest."

https://towardsdatascience.com/why-llms-arent-a-one-size-fits-all-solution-for-enterprises/

#AI #GenerativeAI #LLMs #RFMs #StructuredData #Databases #Graphs

"The culmination of this research and foresight led my team and me to create the first Relational Foundation Model (RFM) for business data. Its purpose is to enable machines to reason directly over structured data, to understand how entities, such as customers, transactions, and products, connect. By knowing the relationships between these entities, we then enable users to make accurate predictions from those specific relationships and patterns.

Unlike LLMs, RFMs have been designed for structured relational data. RFMs are pretrained on a number of (synthetic) datasets as well as on a number of tasks over structured business data. Like LLMs, RFMs can be simply prompted to produce instant responses to a wide variety of predictive tasks over a given database, all without task-specific or database-specific training.

We wanted a system that could learn directly from how real databases are structured, and without all the usual manual setup. To make that possible, we treated each database like a graph: tables became node types, rows turned into nodes, and foreign keys linked everything together. This way, the model could actually “see” how things like customers, transactions, and products connect and change over time.

At the heart of it, the model combines a column encoder with a relational graph transformer. Every cell in a table is turned into a small numerical embedding based on what kind of data it holds, whether it’s a number, category, or a timestamp. The Transformer then looks across the graph to pull context from related tables, which helps the model adapt to new database schemas and data types.

For users to input which predictions they’d like to make, we built a simple interface called Predictive Query Language (PQL). It lets users describe what they want to predict, and the model takes care of the rest."

https://towardsdatascience.com/why-llms-arent-a-one-size-fits-all-solution-for-enterprises/

#AI #GenerativeAI #LLMs #RFMs #StructuredData #Databases #Graphs