Leogrin commited on
Commit
45a2270
·
verified ·
1 Parent(s): cd12352

Add v3 OOD checkpoints + README

Browse files

Adds tabpfn-v3-{classifier,regressor}-v3_ood.ckpt.

Built from the v3 release defaults (identical state_dict) with only PREPROCESS_TRANSFORMS swapped to OOD-robust recipes:
- classifier: squashing_scaler_max10 + none (200 features)
- regressor: quantile_uni_extrapolate + squashing_scaler_max10 (500 features)

All other inference_config fields inherited from the defaults unchanged. The regressor depends on the quantile_uni_extrapolate preset added in PriorLabs/TabPFN#971 (merged); a TabPFN release that contains that merge is required to load it.

README.md CHANGED
@@ -26,7 +26,7 @@ tags:
26
  ### Model Overview
27
  TabPFN-3 is a transformer-based foundation model that uses in-context-learning to solve tabular prediction problems in a forward pass.
28
  Inference code can be found at [https://github.com/PriorLabs/TabPFN](https://github.com/PriorLabs/TabPFN).
29
- More details can be found in the [Model Report](https://arxiv.org/pdf/2605.13986).
30
 
31
  ### Getting started
32
  First, install the inference package:
@@ -65,6 +65,8 @@ The following specialized checkpoints are available:
65
  | [`tabpfn-v3-classifier-v3_20260417_multiclass.ckpt`](https://huggingface.co/Prior-Labs/tabpfn_3/blob/main/tabpfn-v3-classifier-v3_20260417_multiclass.ckpt) | Classification | Specialized for multiclass classification for datasets with <200k rows |
66
  | [`tabpfn-v3-regressor-v3_20260417_mediumdata.ckpt`](https://huggingface.co/Prior-Labs/tabpfn_3/blob/main/tabpfn-v3-regressor-v3_20260417_mediumdata.ckpt) | Regression | Specialized for regression for datasets with <100k rows and with alternative preprocessing |
67
  | [`tabpfn-v3-regressor-v3_20260506_timeseries.ckpt`](https://huggingface.co/Prior-Labs/tabpfn_3/blob/main/tabpfn-v3-regressor-v3_20260506_timeseries.ckpt) | Regression / Time-series forecasting | Fine-tuned on synthetic time-series data; used by default in TabPFN-TS-3 |
 
 
68
 
69
  To use one of these checkpoints, pass its filename via `model_path`:
70
 
@@ -115,12 +117,10 @@ v1.0: initial release.
115
 
116
  ### Citation
117
  ```
118
- @misc{grinsztajn2026tabpfn3technicalreport,
119
- title={TabPFN-3: Technical Report},
120
- author={Léo Grinsztajn and Klemens Flöge and Oscar Key and Felix Birkel and Philipp Jund and Brendan Roof and Mihir Manium and Shi Bin and Hoo and Magnus Bühler and Anurag Garg and Dominik Safaric and Jake Robertson and Benjamin Jäger and Simone Alessi and Adrian Hayler and Vladyslav Moroshan and Lennart Purucker and Philipp Singer and Alan Arazi and Julien Siems and Jan Hendrik Metzen and Georg Grab and Nick Erickson and Siyuan Guo and Eliott Kalfon and Simon Bing and David Salinas and Clara Cornu and Lilly Charlotte Wehrhahn and Diana Kriuchkova and Kursat Kaya and Lydia Sidhoum and Marie Salmon and Jerry Chen and Madelon Hulsebos and Yann LeCun and Samuel Müller and Bernhard Schölkopf and Sauraj Gambhir and Noah Hollmann and Frank Hutter},
121
  year={2026},
122
- eprint={2605.13986},
123
- archivePrefix={arXiv},
124
- url={https://arxiv.org/abs/2605.13986},
125
  }
126
  ```
 
26
  ### Model Overview
27
  TabPFN-3 is a transformer-based foundation model that uses in-context-learning to solve tabular prediction problems in a forward pass.
28
  Inference code can be found at [https://github.com/PriorLabs/TabPFN](https://github.com/PriorLabs/TabPFN).
29
+ More details can be found in the [Model Report](https://storage.googleapis.com/prior-labs-tabpfn-public/reports/TabPFN_3_model_report.pdf).
30
 
31
  ### Getting started
32
  First, install the inference package:
 
65
  | [`tabpfn-v3-classifier-v3_20260417_multiclass.ckpt`](https://huggingface.co/Prior-Labs/tabpfn_3/blob/main/tabpfn-v3-classifier-v3_20260417_multiclass.ckpt) | Classification | Specialized for multiclass classification for datasets with <200k rows |
66
  | [`tabpfn-v3-regressor-v3_20260417_mediumdata.ckpt`](https://huggingface.co/Prior-Labs/tabpfn_3/blob/main/tabpfn-v3-regressor-v3_20260417_mediumdata.ckpt) | Regression | Specialized for regression for datasets with <100k rows and with alternative preprocessing |
67
  | [`tabpfn-v3-regressor-v3_20260506_timeseries.ckpt`](https://huggingface.co/Prior-Labs/tabpfn_3/blob/main/tabpfn-v3-regressor-v3_20260506_timeseries.ckpt) | Regression / Time-series forecasting | Fine-tuned on synthetic time-series data; used by default in TabPFN-TS-3 |
68
+ | [`tabpfn-v3-classifier-v3_ood.ckpt`](https://huggingface.co/Prior-Labs/tabpfn_3/blob/main/tabpfn-v3-classifier-v3_ood.ckpt) | Classification | Same weights as the default classifier but with OOD-robust preprocessors bundled (`squashing_scaler_max10` + `none`). Useful when test inputs may fall outside the training distribution. |
69
+ | [`tabpfn-v3-regressor-v3_ood.ckpt`](https://huggingface.co/Prior-Labs/tabpfn_3/blob/main/tabpfn-v3-regressor-v3_ood.ckpt) | Regression | Same weights as the default regressor but with OOD-robust preprocessors bundled (`quantile_uni_extrapolate` + `squashing_scaler_max10`). Linearly extrapolates past the training range instead of clamping. Requires `tabpfn` from the public main branch (post-#971 merge). |
70
 
71
  To use one of these checkpoints, pass its filename via `model_path`:
72
 
 
117
 
118
  ### Citation
119
  ```
120
+ @misc{TabPFN3,
121
+ title={TabPFN-3},
122
+ author={Prior Labs},
123
  year={2026},
124
+ note={Coming soon}
 
 
125
  }
126
  ```
tabpfn-v3-classifier-v3_ood.ckpt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f973d994ece09f9e28e9319d644bcff3cb0bec9c3170e67e14e71f53df92d18b
3
+ size 212815699
tabpfn-v3-regressor-v3_ood.ckpt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:fe1a5aa90cdd7f06e8736a68015b1bb570a413bd302de2d1038364272a8c24d6
3
+ size 233300137