Show HN: nanoTabPFN – A Lightweight and Educational Reimplementation of TabPFN (github.com)

nanoTabPFN is a compact, educational reimplementation of TabPFN that you can clone, install with a small Python stack (numpy, torch, schedulefree, h5py, scikit-learn, openml, seaborn) and pretrain in minutes. The repo keeps core code under 500 lines: model.py (~200 LOC) exposes a sklearn-like NanoTabPFNClassifier, train.py (~200 LOC) provides a simple training loop and PriorDumpDataLoader, and experiment.ipynb reproduces paper experiments. The author-provided prior dump (300k_150x5_2.h5) lets you pretrain quickly; the default example trains for 2,500 steps with batch_size 32 and lr=4e-3. The example NanoTabPFNModel uses small transformer-like settings (embedding_size=96, num_attention_heads=4, mlp_hidden_size=192, num_layers=3, num_outputs=2) and supports fit/predict/predict_proba on standard datasets. Significance: nanoTabPFN demystifies TabPFN-style priors and transformer architectures for tabular data by making the full pipeline transparent, reproducible and hackable. For researchers and students it’s a low-friction sandbox to study how synthetic priors, architecture choices, and training regimes affect performance, and to iterate on extensions (regression, alternate priors, ensembling). The project also points to a larger TFM-Playground for more features. A formal preprint is available (arXiv:2511.03634) for those who want the complete technical background.