Skip to main content
TabPFN accepts up to 20,000 columns, with 6,000 recommended, and handles wide tables without a selection step. See Models for the limits of older versions. Feature selection is no longer a routine requirement. It still helps in two cases: tables with many columns that carry no signal, and tables beyond the recommended width where you want to control which columns the model sees.

​
How TabPFN covers wide tables

Each estimator sees up to 768 features. When a table is wider, TabPFN distributes columns across estimators so that every column is seen by some estimator, and averages the estimators’ predictions. With the default 8 estimators this covers about 6,000 columns. Beyond that, raise n_estimators so more columns are covered, or select features yourself.
See Wide tables and feature selection for the feature subsampling methods and their settings.

​
When to select features

  • Many noisy columns. Irrelevant features dilute attention. If you know that most columns carry no signal, filtering them helps both accuracy and speed.
  • Beyond the recommended width. Above about 6,000 columns, either raise n_estimators or select features so each estimator sees the columns that matter.
  • Latency budgets. Fewer columns mean faster fits and predictions, and a smaller KV cache.

​
Approaches

Greedy feature selection removes features individually and checks performance. This works particularly well on smaller data with low computational cost. Mutual information filtering ranks features by mutual information with the target and keeps the top k:
Importance-based subsampling inside TabPFN keeps the most important columns per estimator without a separate step. Set FEATURE_SUBSAMPLING_METHOD to "gini_feature_importance":
PCA or TruncatedSVD reduces dimensionality while retaining variance. Use it as a last resort: it mixes columns, so the model loses the per-column structure it is built to read. Validate every selection step against the unchanged table on held-out data. The raw table is often the strongest baseline.