File size: 2,975 Bytes
6e4d18e
 
 
 
9a448d6
 
 
 
 
 
 
6e4d18e
9a448d6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
---
license: other
license_name: tabpfn-3-license-v1.0
license_link: LICENSE
pipeline_tag: tabular-regression
tags:
- tabular
- regression
- agriculture
- maize-yield
- tabpfn-3
---

### Model Overview
`tabpfn-v3-regressor-cn_finetuned_full.ckpt` is the **full training-snapshot** of the
fine-tuned derivative of [TabPFN-v3](https://github.com/PriorLabs/TabPFN)
(`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
```python
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`](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.
```