Introduction

Code for AISTATS2025 paper: Prior-Fitted Networks Scale to Larger Datasets When Treated as Weak Learners.

Initial code for submission. Further revision coming soon.

Requirements

See requirements.txt which is from previous work of TabPFN v1.

Reproduction of BoostPFN in Table 3

To run small datasets lower than 5000:

python main_10times.py --modelname gboost_tabpfnV2 --updating exphadamard

To run large datasets:

python largedataset_boostpfn.py --modelname gboost_tabpfnV2 --gpu 0 --sampling_size 0.001 --test_batch 50000 --seed 5 --maxsample 500 --updating exphadamard --endnum 0 --step 100 --startnum -1 --ensemble_num 10

Main Arguments:

--updating: the updating method for boostpfn, can choose from exphadamard,hadamard,adaboost

--ensemble_num: the number of boosting rounds