TabFM
A zero-shot foundation model for tabular data that performs tabular prediction tasks without fine-tuning on specific datasets. Building on the research direction established by TabPFN, it supports cross-domain generalization and in-context learning capabilities.
Timeline (last 90 days)
In the demo, TabFM runs locally on a DGX Spark, with data stored in a Neon-based Serverless Postgres.
Google Research released TabFM, a foundation model built specifically for tabular data.
TabFM does not require users to train with their own datasets; instead, it takes labeled rows and unlabeled rows to be predicted as joint input and directly infers answers in a single forward pass.
TabFM is trained entirely on synthetic data, generating row data by sampling random causal graphs for each table and forcing the model to predict masked cells.
TabFM's weights are currently restricted to non-commercial research use only.
TabFM is trained on synthetic causal graphs and has never seen any column names or labels during training, and therefore cannot understand the semantics of columns.
The researchers proposed an architecture called TabFM Auto, which keeps TabFM frozen as the inference engine while an LLM-driven agent takes over pipeline tasks such as inspecting column names, writing data transformations, selecting in-context rows, and calibrating predictions.
Under the condition of a single table and zero feature engineering, TabFM outperformed the author's previously carefully tuned tree model in a customer churn prediction demo, but the score was only 0.60.
Compared to Google's TabFM (which is approximately 30 times larger than TabPFN 3.5), TabPFN 3.5 achieves 20× faster forward passes at equivalent quality, or scores more than 100 ELO points higher while also being faster.
All Facts (9)
In the demo, TabFM runs locally on a DGX Spark, with data stored in a Neon-based Serverless Postgres.
50%UnverifiedThe researchers proposed an architecture called TabFM Auto, which keeps TabFM frozen as the inference engine while an LLM-driven agent takes over pipeline tasks such as inspecting column names, writing data transformations, selecting in-context rows, and calibrating predictions.
50%UnverifiedUnder the condition of a single table and zero feature engineering, TabFM outperformed the author's previously carefully tuned tree model in a customer churn prediction demo, but the score was only 0.60.
50%UnverifiedTabFM does not require users to train with their own datasets; instead, it takes labeled rows and unlabeled rows to be predicted as joint input and directly infers answers in a single forward pass.
50%UnverifiedTabFM is trained entirely on synthetic data, generating row data by sampling random causal graphs for each table and forcing the model to predict masked cells.
50%UnverifiedTabFM's weights are currently restricted to non-commercial research use only.
50%UnverifiedTabFM is trained on synthetic causal graphs and has never seen any column names or labels during training, and therefore cannot understand the semantics of columns.
50%UnverifiedCompared to Google's TabFM (which is approximately 30 times larger than TabPFN 3.5), TabPFN 3.5 achieves 20× faster forward passes at equivalent quality, or scores more than 100 ELO points higher while also being faster.
50%UnverifiedGoogle Research released TabFM, a foundation model built specifically for tabular data.
50%