This repository hosts a fitted TabPFNv2 classifier and the full, reproducible benchmark that establishes the result. TabPFNv2 is a prior-data fitted network: a transformer pre-trained on millions of synthetic tabular tasks that performs classification in a single forward pass (in-context learning).

🥇 SOTA Benchmark : ranked by ROC-AUC

Dataset: UCI Breast Cancer Wisconsin Diagnostic = 569 rows, 30 numerical features, binary (212 malignant / 357 benign). Protocol: RepeatedStratifiedKFold, seed 42, single fixed config per model, no tuning on test.

Rank	Model	ROC-AUC	Accuracy	CV Folds	Time (s)
🥇 1	TabPFNv2 (this model, CPU)	0.9970 ± 0.0039	0.9807 ± 0.0116	5×1	214.8
🥈 2	Logistic Regression (standardized)	0.9947 ± 0.0077	0.9778 ± 0.0169	10×3	10.9
🥉 3	CatBoost	0.9939 ± 0.0084	0.9707 ± 0.0228	10×3	91.0
4	LightGBM	0.9934 ± 0.0084	0.9672 ± 0.0206	10×3	60.6
5	XGBoost	0.9933 ± 0.0088	0.9661 ± 0.0256	10×3	60.1
6	HistGradientBoosting	0.9919 ± 0.0107	0.9608 ± 0.0278	10×3	70.9
7	RandomForest	0.9905 ± 0.0129	0.9596 ± 0.0276	10×3	75.4

TabPFNv2 is the single best model on both ROC-AUC and Accuracy, and has the lowest variance (±0.0039 AUC) ; the most stable estimate of all. Published SOTA on WDBC under cross-validation is ~97–98.6% accuracy / ~0.99 AUC, so this result sits at the ceiling of the state of the art.

Benchmark context

Grinsztajn, Oyallon & Varoquaux, "Why do tree-based models still outperform deep learning on tabular data?", NeurIPS 2022 (arXiv 2207.08815, Jul 2022) — source of the tuned GBDT recipes and the AUC/accuracy CV evaluation protocol.
TabArena (Erickson et al.), "A Living Benchmark for Machine Learning on Tabular Data", Jun 2025 (arXiv 2506.16791; NeurIPS 2025 Spotlight) — living leaderboard where the TabPFN family ranks #2–#3 globally and #1 on small datasets (≤10k rows), the regime WDBC falls in. CatBoost/LightGBM/XGBoost are the strongest CPU tree models there.

💻 Library versions used

Library	Version
tabpfn	2.0.9
scikit-learn	1.6.1
xgboost	3.2.0
lightgbm	4.6.0
catboost	1.2.10
Python	3.12

🚀 Usage

python

import pickle
from huggingface_hub import hf_hub_download
from sklearn.datasets import load_breast_cancer
from sklearn.model_selection import train_test_split
from sklearn.metrics import roc_auc_score, accuracy_score

# Load the fitted model
path = hf_hub_download("AurelPx/tabpfnv2-ultimate-breast-cancer-detection", "tabpfnv2_wdbc.pkl")
clf = pickle.load(open(path, "rb"))

# Evaluate on the same holdout split (seed 42)
X, y = load_breast_cancer(return_X_y=True)
X = X.astype("float32")
_, Xte, _, yte = train_test_split(X, y, test_size=0.25, random_state=42, stratify=y)

proba = clf.predict_proba(Xte)[:, 1]
print("AUC", roc_auc_score(yte, proba))       # ~0.998
print("ACC", accuracy_score(yte, proba > 0.5)) # ~0.972

Install: pip install tabpfn==2.0.9 scikit-learn huggingface_hub. On first use TabPFNv2 weights are fetched from the Hub (no license token required for 2.0.9). For CPU, the script sets TABPFN_ALLOW_CPU_LARGE_DATASET=1 and device="cpu".

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "AurelPx/tabpfnv2-ultimate-breast-cancer-detection"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id)

For non-causal architectures, replace AutoModelForCausalLM with the appropriate AutoModel class.