TabDPT Checkpoints

Pre-trained TabDPT model weights trained with three different random seeds. Each checkpoint is from epoch 2040 of production training.

Files

File	Training Seed
production_seed42.safetensors	42
production_seed123.safetensors	123
production_seed456.safetensors	456

Model Architecture

• Embedding size: 512
• Attention heads: 8
• Layers: 12
• Hidden factor: 2
• Max features: 100
• Max classes: 10

Benchmark Results

Classification: Breast Cancer (binary, 30 features)

Checkpoint	Accuracy	Ensemble Accuracy
seed42	99.4%	99.4%
seed123	98.8%	98.8%
seed456	98.2%	98.2%
HF default (Layer6/TabDPT)	—	99.4%

Classification: Wine (3-class, 13 features)

Checkpoint	Accuracy	Ensemble Accuracy
seed42	100%	100%
seed123	100%	100%
seed456	100%	100%
HF default (Layer6/TabDPT)	—	100%

Regression: Diabetes (10 features)

Checkpoint	MSE	Correlation
seed42	2618.6	0.718
seed123	2655.1	0.713
seed456	2795.5	0.701
HF default (Layer6/TabDPT)	2673.1	0.711

Training Stats (from checkpoint metadata)

Metric	seed42	seed123	seed456
CC18 Accuracy	0.877	0.878	0.879
CC18 F1	0.870	0.872	0.873
CC18 AUC	0.927	0.927	0.928
CTR Correlation	0.830	0.830	0.827
CTR R²	0.726	0.730	0.725

Format

These checkpoints were converted from PyTorch Lightning .ckpt files (which include optimizer state, ~295MB each) to SafeTensors format (model weights only, ~103MB each). This is the same format used by the official Layer6/TabDPT release. The tabdpt package natively loads SafeTensors via the model_weight_path argument — no extra conversion needed.

Usage

from tabdpt import TabDPTClassifier
from huggingface_hub import hf_hub_download

# Download once (cached afterwards)
path = hf_hub_download("dwahdany/TabDPT", "production_seed42.safetensors")

# Use exactly like the default model
clf = TabDPTClassifier(model_weight_path=path)
clf.fit(X_train, y_train)
preds = clf.predict(X_test)

Works identically with TabDPTRegressor.