From 91b99157426104febbb69caa1fdef01a7d61fe28 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?=E2=80=9Cmohansree14=E2=80=9D?=
 <“mohansree1709@gmail.com”>
Date: Thu, 20 Aug 2026 22:38:56 +0100
Subject: [PATCH] docs(auroc): clarify per-batch logit/probability check in
 BinaryAUROC

Fixes #2195 - the sigmoid-vs-probability check in update() runs per
batch, not across the full dataset, which can silently produce
inconsistent conversion behavior across batches. Documented this in
the class docstring.
---
 src/torchmetrics/classification/auroc.py | 7 +++++++
 1 file changed, 7 insertions(+)

diff --git a/src/torchmetrics/classification/auroc.py b/src/torchmetrics/classification/auroc.py
index 68960e203e9..473ce787fbf 100644
--- a/src/torchmetrics/classification/auroc.py
+++ b/src/torchmetrics/classification/auroc.py
@@ -57,6 +57,13 @@ class BinaryAUROC(BinaryPrecisionRecallCurve):
       therefore only contain {0,1} values (except if `ignore_index` is specified). The value 1 always encodes the
       positive class.
 
+    .. note::
+       The probabilities-or-logits check above is applied independently on every ``update()`` call, i.e. per batch
+       rather than across the full accumulated dataset. If you call ``update()`` with small batches, a batch whose
+       predictions happen to all fall inside ``[0, 1]`` will **not** have sigmoid applied, even if other batches in
+       the same run were treated as logits. To avoid inconsistent results, either pass probabilities (post-sigmoid)
+       consistently across all batches, or apply ``torch.sigmoid`` yourself before calling ``update()``.
+
     As output to ``forward`` and ``compute`` the metric returns the following output:
 
     - ``b_auroc`` (:class:`~torch.Tensor`): A single scalar with the auroc score.