From 27f36fbba3dcc825aa81ca0593d476ff1f81ce90 Mon Sep 17 00:00:00 2001
From: Philip Hartout <philip.hartout@protonmail.com>
Date: Thu, 26 Mar 2026 10:32:49 +0100
Subject: [PATCH] Use TabPFN v2.5 weights as default classifier

Switch default_classifier() to use TabPFNClassifier.create_default_for_version
with ModelVersion.V2_5 instead of the plain constructor (which uses v2.0).
Update README to reflect the new default and document the license differences
between v2.0 (Apache 2.0 + attribution, commercial OK) and v2.5
(non-commercial). Add logistic regression as a lightweight alternative.
---
 README.md                                      | 18 ++++++++++++------
 polygraph/metrics/base/polygraphdiscrepancy.py |  9 +++++++--
 2 files changed, 19 insertions(+), 8 deletions(-)

diff --git a/README.md b/README.md
index d4d38ef..7543c17 100644
--- a/README.md
+++ b/README.md
@@ -142,24 +142,30 @@ print(pgd.compute(generated)) # {'pgd': ..., 'pgd_descriptor': ..., 'subscores':
 
 `pgd_descriptor` provides the best descriptor used to report the final score.
 
-By default, PGD uses TabPFN v2 weights. To use TabPFN v2.5 weights instead, pass a custom classifier. The v2.5 weights are hosted on a gated Hugging Face repository ([Prior-Labs/tabpfn_2_5](https://huggingface.co/Prior-Labs/tabpfn_2_5)) and require authentication:
+By default, PGD uses TabPFN v2.5 weights. The v2.5 weights are hosted on a gated Hugging Face repository ([Prior-Labs/tabpfn_2_5](https://huggingface.co/Prior-Labs/tabpfn_2_5)) and require authentication:
 
 ```bash
 pip install huggingface_hub
 huggingface-cli login
 ```
 
-Then:
+Alternatively, you can use TabPFN v2.0 weights, which are licensed under the Prior Labs License (Apache 2.0 with an additional attribution clause) and permit commercial use. The v2.5 weights, in contrast, use a non-commercial license that prohibits commercial and production use without a separate enterprise license from Prior Labs:
 
 ```python
 from tabpfn import TabPFNClassifier
 from polygraph.metrics import StandardPGD
 
-classifier = TabPFNClassifier.create_default_for_version(
-    "v2.5", device="auto", n_estimators=4
-)
+classifier = TabPFNClassifier(device="auto", n_estimators=4)
 pgd = StandardPGD(reference, classifier=classifier)
-print(pgd.compute(generated))
+```
+
+A logistic regression classifier can also be used as a lightweight alternative, although it yields a looser bound in practice:
+
+```python
+from sklearn.linear_model import LogisticRegression
+from polygraph.metrics import StandardPGD
+
+pgd = StandardPGD(reference, classifier=LogisticRegression())
 ```
 
 #### Validity, uniqueness and novelty
diff --git a/polygraph/metrics/base/polygraphdiscrepancy.py b/polygraph/metrics/base/polygraphdiscrepancy.py
index f4d8c48..a4ee81c 100644
--- a/polygraph/metrics/base/polygraphdiscrepancy.py
+++ b/polygraph/metrics/base/polygraphdiscrepancy.py
@@ -62,6 +62,7 @@
 from sklearn.preprocessing import StandardScaler
 from packaging.version import Version
 from tabpfn import TabPFNClassifier
+from tabpfn.classifier import ModelVersion
 
 from polygraph import GraphType
 from polygraph.metrics.base.interface import GenerationMetric
@@ -107,7 +108,7 @@ def default_classifier() -> TabPFNClassifier:
     """Create the default TabPFN classifier used by PGD.
 
     Returns:
-        A TabPFNClassifier with default settings (auto device, 4
+        A TabPFNClassifier with v2.5 weights (auto device, 4
         estimators). Requires ``tabpfn >= 2.0.9``.
     """
     tabpfn_ver = Version(version("tabpfn"))
@@ -116,7 +117,11 @@ def default_classifier() -> TabPFNClassifier:
             "TabPFN >= 2.0.9 is required. "
             "Install with `pip install 'tabpfn>=2.0.9'`."
         )
-    return TabPFNClassifier(device="auto", n_estimators=4)
+    return TabPFNClassifier.create_default_for_version(
+        ModelVersion.V2_5,
+        device="auto",
+        n_estimators=4,
+    )
 
 
 class PolyGraphDiscrepancyResult(TypedDict):