rouluan/Agro-TabPFN

Model Overview

tabpfn-v3-regressor-cn_finetuned_full.ckpt is the full training-snapshot of the fine-tuned derivative of TabPFN-v3 (tabpfn-v3-regressor-v3_default), fine-tuned on China provincial maize yield data (new dataset: feature_dataset.csv).

This is the resumable checkpoint. Besides the fine-tuned state_dict it also contains the AdamW optimizer state (exp_avg / exp_avg_sq), epoch and loss metadata — i.e. the full ~444 MB of optimizer buffers that make a 668 MB file. Use this file to continue fine-tuning from the best epoch.

Fine-tuning summary

• Data: feature_dataset.csv, years 2003–2022, 602 rows, 31 provinces
• Split: train 2003–2016 (422) / val 2017–2019 (90) / test 2020–2022 (90); 96 features
• Best val MSE 0.2716 at epoch 26 of 30 (early-stop patience 5)
• lr 3e-6, wd 0.01, MSE 1.0 / CRPS 1.0 / MAE 0.3, context 256, seed 42

Internal test (2020–2022, 90 rows, 16-ensemble) vs base model

Metric	Base	CN-finetuned
RMSE	0.6865	0.6207
R²	0.6308	0.6981
MAE	0.4592	0.4494
Pearson r	0.8681	0.8914
KGE	0.8507	0.8554

Usage for inference

from tabpfn import TabPFNRegressor
reg = TabPFNRegressor(model_path="tabpfn-v3-regressor-cn_finetuned_full.ckpt",
                      device="cuda", n_estimators=16, random_state=42)
reg.fit(X_train, y_train)
y_pred = reg.predict(X_test)

File integrity

• Full snapshot ckpt SHA-256: 53119ca48958a7b7eb5ca06752eecc2348f0b2ffa58d45d4f5ae79f02793f630

Licensing

Model weights are released under the TABPFN-3 Non-Commercial License v1.0. The full license text is included in LICENSE and must be passed on to any third party who receives this model or a derivative of it.

This model is a modified derivative. tabpfn-v3-regressor-cn_finetuned_full.ckpt is a fine-tuned version of tabpfn-v3-regressor-v3_default. It is not an official product of Prior Labs GmbH and has not been endorsed, approved, or validated by the Company.

Research and internal evaluation are allowed; any commercial or production use is prohibited. For commercial use contact sales@priorlabs.ai.

Per §3.d of the license, this model and its derivatives must not be distributed, hosted, or made available as part of a hosted, managed, API, or SaaS service.

Required attribution notice:

The TABPFN-3 Model is licensed by Prior Labs GmbH under the TABPFN-3 Non-Commercial License.
Copyright © Prior Labs GmbH 2026.
THE SERVICES ARE PROVIDED FREE OF CHARGE: COMPANY SHALL NOT BE LIABLE FOR DAMAGES RESULTING FROM SLIGHT NEGLIGENCE. LIABILITY FOR GROSS NEGLIGENCE AND INTENTIONAL MISCONDUCT REMAINS UNAFFECTED.