PriorLabs/TabPFN

⚡ TabPFN: Foundation Model for Tabular Data ⚡

What it solves

TabPFN provides a high-performance, transformer-based approach to tabular machine learning. It is designed to solve small-to-medium-sized classification and regression problems with high accuracy and minimal setup, bypassing the need for extensive data preprocessing or complex hyperparameter tuning.

How it works

TabPFN is a "Prior-Data Fitted Network"—a foundation model trained on a massive variety of synthetic tabular datasets. By learning the underlying patterns of tabular data during pre-training, it can perform inference on new, unseen datasets in a single forward pass. It uses a transformer architecture with row-wise and cross-row attention mechanisms to process tabular inputs directly, effectively treating the dataset as a sequence of tokens.

Who it’s for

It is ideal for data scientists and researchers working with tabular data who want a "plug-and-play" model that performs competitively with state-of-the-art gradient boosting methods. It is particularly useful for rapid prototyping and scenarios where data preprocessing (like scaling or encoding) is undesirable.

Highlights

  • Zero Preprocessing: Works directly on raw data without requiring one-hot encoding or feature scaling.
  • Fast Inference: Designed for rapid predictions, often outperforming traditional models on smaller datasets.
  • Transformer Architecture: Uses advanced attention mechanisms to capture complex relationships between rows and features.
  • Flexible Ecosystem: Includes extensions for interpretability (SHAP), outlier detection, and embedding extraction.
  • Hardware Optimized: Supports GPU acceleration via PyTorch/CUDA and Apple Silicon, with a fallback for CPU usage.

Related

  • Project
    PriorLabs/tabpfn-time-series

    TabPFN‑TS is a Python library that turns univariate time‑series forecasting into a tabular regression problem and uses the pretrained TabPFN foundation model (default TabPFN‑3.5) to produce zero‑shot point and probabilistic forecasts. It works with known‑future covariates, runs fast via a local CPU or a cloud client (no GPU needed), and requires only a pip install. Limitations include lack of past dynamic or static covariates and handling multivariate targets as independent series.

  • Project
    LAMDA-Tabular/TALENT

    TALENT is an open‑source benchmark and toolbox for tabular machine‑learning. It bundles >30 deep‑learning and classical methods (including the latest foundation models like TabPFN‑3.5, LimiX, NVIDIA Kumo), provides 300 ready‑to‑use datasets, unified preprocessing, hyper‑parameter tuning, and evaluation (accuracy, AUC, calibration). Install via pip, call `get_method("ft_transformer")` or any other model, and run with a single `fit`/`predict` call. The project is released under BSD‑3‑Clause with separate licenses for some pretrained weights.

  • Project
    PriorLabs/tabpfn-extensions

    An experimental add‑on library that expands the TabPFN tabular‑ML model with interpretability, multi‑class handling, image columns, unsupervised utilities, embeddings, Bayesian optimisation, and statistical testing. Install via pip (pointing at the Git repo), choose a local or cloud TabPFN backend, and use the bundled notebooks or per‑extension READMEs for guidance.

  • Project
    NVIDIA/TransformerEngine

    NVIDIA Transformer Engine is an open‑source library that speeds up Transformer model training on NVIDIA GPUs by providing fused kernels and automatic mixed‑precision APIs (FP8, MXFP8, NVFP4, BF16, FP16) for PyTorch and JAX. It handles scaling‑factor management, quantization metadata, and distributed communication, letting users adopt low‑precision training without changing their code. Integrated with major LLM frameworks (Hugging Face, Lightning, Megatron‑LM, etc.), it targets large‑scale language, vision, and MoE models on Hopper/Blackwell/Rubin hardware.

  • Project
    PaddlePaddle/PaddleFormers

    PaddleFormers is a PaddlePaddle‑based Transformers library that mirrors Hugging Face’s API while adding high‑performance distributed training (tensor, pipeline, expert parallelism, FP8, FlashMLA, etc.). It supports 100+ LLM/VLM families (DeepSeek, GLM‑4.5, Qwen, Llama, ERNIE‑4.5, Gemma‑3, etc.), full training pipelines (pre‑training, SFT/LoRA, DPO, function‑call, reasoning), and native support for Chinese domestic GPUs. Install via Docker or pip, then use familiar `AutoTokenizer`/`AutoModelForCausalLM` classes for inference or the `paddleformers-cli` for large‑scale training.