Skip to main content
The TabPFN Unsupervised Extension brings TabPFN’s foundation-model reasoning to unsupervised anomaly detection. TabPFN estimates the likelihood of each sample under its learned distribution. Samples with low joint probability are considered anomalous or outliers. Outlier detection example

​
Getting Started

First, install the unsupervised extension: 
pip install "tabpfn-extensions[unsupervised]"
Then, load your data, initialize a TabPFN classifier and regressor model, and predict outlier scores with the TabPFNUnsupervisedModel class: 
import torch
from sklearn.datasets import load_breast_cancer
from tabpfn import TabPFNClassifier, TabPFNRegressor
from tabpfn_extensions.unsupervised import TabPFNUnsupervisedModel

# Load dataset
df = load_breast_cancer(return_X_y=False)
X = df["data"]
attribute_names = df["feature_names"]

# Initialize models
clf = TabPFNClassifier(n_estimators=4)
reg = TabPFNRegressor(n_estimators=4)
model_unsupervised = TabPFNUnsupervisedModel(
    tabpfn_clf=clf, tabpfn_reg=reg
)

# Compute anomaly scores
model_unsupervised.fit(X)

scores = model_unsupervised.outliers(
    torch.tensor(X, dtype=torch.float32),
    n_permutations=10,
)

print("Outlier scores:", scores[:10])

​
How It Works

TabPFN decomposes the joint feature probability using the chain rule: P(X)=∏i=1dP(Xi∣X<i)P(X) = \prod_{i=1}^{d} P(X_i \mid X_{<i}) Each conditional term is predicted using a TabPFN model:
  • Continuous features: via probability density (PDF)
  • Categorical features: via probability mass function (PMF)
The model sums log-probabilities across features to produce a sample-level log-likelihood score. Multiple random permutations are averaged to ensure stability across different feature orderings.

Google Colab Example

Check out our Google Colab for a demo.