What is a Data Language Model?
A Data Language Model reads tabular data the way a language model reads text. You hand it tables and databases in their original form, and it tells you what the data is, how the pieces relate, what is missing, and what can be predicted.
The problem with tables
Most of the world's data sits in tables, and most of the work on it is preparation: matching columns by hand, inventing join keys, cleaning, labelling. Language models skipped that step for text. They learned from raw text and handled the rest at inference time. A Data Language Model applies the same idea to rows and columns.
What goes in
- Tables and databases as they are, from any number of sources.
- No shared keys. The sources do not have to agree on an identifier.
- No labels and no preprocessing. Column names can be coded or missing.
What comes back
One call returns one bundle: the sector the data belongs to, a profile of every column, a map between tables, filled-in missing values, and predictions for a chosen target. Every row-level result carries a confidence. The next article walks through each part.
The category
I wrote the paper that defines the category: arXiv 2605.06290. A shared definition gives the field one target to build and measure against.
How it relates to tabular foundation models
A Data Language Model is a tabular foundation model with a stricter contract: it reads the raw table itself, with no serialization to text, no preprocessing, and no dependence on column names. Scaled up, it becomes a large tabular foundation model that carries prediction, imputation, anomaly detection, and matching in one checkpoint. See the architecture, the benchmark stories, and the guide to AI and machine learning for tabular data.
Frequently asked questions
What is a Data Language Model?
A foundation model that reads raw tables directly, the way a language model reads text, and returns understanding, relationships, missing-value fills, and predictions in one pass.
Is a Data Language Model the same as a large language model?
No. A large language model reads text tokens. A Data Language Model reads table cells and keeps the row and column structure.
Where is the paper?
On arXiv: 2605.06290.