tabpfn_2_5

Model Overview

TabPFN-2.5 is a transformer-based foundation model that uses in-context-learning to solve tabular prediction problems in a forward pass. Inference code can be found at https://github.com/PriorLabs/tabPFN.

Getting started

First, install the inference package:

pip install tabpfn

Fitting a classifier and predicting looks like this:

from sklearn.datasets import load_breast_cancer
from sklearn.model_selection import train_test_split
from tabpfn import TabPFNClassifier

# Load data
X, y = load_breast_cancer(return_X_y=True)
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.5, random_state=42)

# Initialize a classifier
clf = TabPFNClassifier()  # Uses TabPFN 2.5 weights, finetuned on real data.
clf.fit(X_train, y_train)


# Predict probabilities
prediction_probabilities = clf.predict_proba(X_test)
# Predict labels
predictions = clf.predict(X_test)
print("Accuracy", accuracy_score(y_test, predictions))

For more examples (e.g. how to train a regressor), see the github repo: https://github.com/PriorLabs/tabPFN!

Developers & Affiliations

Developed by Prior Labs.

Intended Use

Regression and classification tasks with ≤50 000 samples and ≤2000 features in structured tabular format.

Not Intended Use

• Not suitable for unstructured data (text, images); use API version for textual features.
• Not tested for >50 000 samples or > 2000 features.

Model Architecture

Transformer with TabPFNv2-like alternating attention with 18-24 layers

Training Data and Priors

• TabPFN-2.5: trained purely on synthetic tabular tasks
• Real-TabPFN-2.5: continued pre-training on real-world datasets (for details please see Appendix C.1 of the model tech report).

Performance Benchmarks

Evaluated on proprietary benchmark collection, TabArena, and RealCause (for a causal version), in each of which it yields new SOTA results by a wide margin. Please see the model tech report for details.

Different checkpoints

Beyond the default checkpoints (tabpfn-v2.5-classifier-v2.5_default.ckpt and tabpfn-v2.5-regressor-v2.5_default.ckpt), the other available checkpoints are experimental and worse on average, and we recommend to always start with the defaults. They can be used as part of an ensembling or hyperparameter optimization system (and are used automatically in AutoTabPFN here) or tried out manually. Their name suffixes refer to what we expect them to be good at.

We add the 🌍 emoji for checkpoints finetuned on real datasets. See the TabPFN-2.5 paper for the list of 43 datasets.

• tabpfn-v2.5-classifier-v2.5_default.ckpt 🌍: default classification checkpoint, finetuned on real-data.
• tabpfn-v2.5-classifier-v2.5_default-2.ckpt: best classification synthetic checkpoint. Use this to get the default TabPFN-2.5 classification model without real-data finetuning.
• tabpfn-v2.5-classifier-v2.5_large-features-L.ckpt: specialized for larger features (up to 500) and small samples (< 5K).
• tabpfn-v2.5-classifier-v2.5_large-features-XL.ckpt: specialized for larger features (up to 1000, could support max_features_per_estimator=1000).
• tabpfn-v2.5-classifier-v2.5_large-samples.ckpt: specialized for larger sample sizes (larger than 30K)
• tabpfn-v2.5-classifier-v2.5_real.ckpt 🌍: other real-data finetuned classification checkpoint. Pretty good overall but bad on large features (>100-200).
• tabpfn-v2.5-classifier-v2.5_real-large-features.ckpt 🌍: other real-data finetuned classification checkpoint, worse on large samples (> 10K)
• tabpfn-v2.5-classifier-v2.5_real-large-samples-and-features.ckpt 🌍: identical to tabpfn-v2.5-classifier-v2.5_default.ckpt
• tabpfn-v2.5-classifier-v2.5_variant.ckpt: pretty good but bad on large features (> 100-200).
• tabpfn-v2.5-regressor-v2.5_default.ckpt: default regression checkpoint, trained on synthetic data only.
• tabpfn-v2.5-regressor-v2.5_low-skew.ckpt: variant specialized at low target skew data (but quite bad on average).
• tabpfn-v2.5-regressor-v2.5_quantiles.ckpt: variant which might be interesting for quantile / distribution estimation, though the default should still be prioritized for this.
• tabpfn-v2.5-regressor-v2.5_real.ckpt 🌍: finetuned on real-data. Best checkpoint among the checkpoints finetuned on real data. For regression we recommend the synthetic-only checkpoint as a default, but this checkpoint is quite a bit better on some datasets.
• tabpfn-v2.5-regressor-v2.5_real-variant.ckpt 🌍: other regression variant finetuned on real data.
• tabpfn-v2.5-regressor-v2.5_small-samples.ckpt: variant slightly better on small (< 3K) samples.
• tabpfn-v2.5-regressor-v2.5_variant.ckpt: other variant, no clear specialty but can be better on a few datasets.

Ethical Considerations

Having been trained purely on synthetic datasets, TabPFN-2.5 is free from dataset leakage from the pretraining stage. However, like for any other tabular prediction method, when applied to high-risk use cases, users should ensure that the labelled data is free of biases. For Real-TabPFN-2.5, you can find the dataset list in Appendix C.1 of the model tech report.

Limitations

Performance can degrade when applied to >50000 data points and/or 2000 features.

Licensing

Model weights released under tabpfn-2.5-license-v1.1.

The license is designed to be permissive for research and limited internal evaluation. It explicitly allows testing, evaluation, and internal benchmarking, so an organization can download the model and run preliminary assessments on its own datasets. The key restriction is that the model, its derivatives, and its outputs cannot be used for any commercial or production purpose. This includes, but is not limited to, revenue-generating products, competitive benchmarking for procurement, client deliverables, or using the model’s results for internal commercial decision-making. For all production use cases, we offer a Commercial Enterprise License. This provides access to our proprietary high-speed inference engine, dedicated support, integration tooling, and other internal models. Please contact us at sales@priorlabs.ai for commercial licensing inquiries.

Version

v1.0: initial release.

Citation

@misc{TabPFN-2.5,\
      title={TabPFN-2.5: Advancing the State of the Art in\
Tabular Foundation Models},\
      author={Léo Grinsztajn and Klemens Flöge and Oscar Key and Felix Birkel and Brendan Roof and Phil Jund and Benjamin Jäger and Adrian Hayler and Dominik Safaric and Simone Alessi, Felix Jablonski and Mihir Manium and Rosen Yu and Anurag Garg and Jake Robertson and Shi Bin (Liam) Hoo and Vladyslav Moroshan and Magnus Bühler and Lennart Purucker and Clara Cornu and Lilly Charlotte Wehrhahn and Alessandro Bonetto and Sauraj Gambhir and Noah Hollmann and Frank Hutter},\
      year={2025}\
}