fix: force test set to full data in load_test_data

--sample_ratio is a train-only starvation knob; the test set is the fixed
generalization ruler and must never be subsampled. Previously load_test_data
inherited self.sample_ratio, which subsampled test too and made test losses
incomparable across a sample_ratio sweep. Force 1.0 unconditionally. Standard
runs use 1.0, so only sub-1.0 train-starved runs are affected.

Author: neko941

strategies/tFL.py

@@ -510,8 +510,12 @@ def load_train_data(
     def load_test_data(
         self, sample_ratio: float = None, shuffle: bool = False
     ) -> DataLoader:
-        if sample_ratio is None:
-            sample_ratio = getattr(self, "sample_ratio", 1.0)
+        # The test set is the fixed generalization ruler and must ALWAYS be full.
+        # --sample_ratio is a train-side starvation knob (e.g. the grokking probe);
+        # subsampling test would make test losses incomparable across a sample_ratio
+        # sweep. Force 1.0 regardless of the passed/instance value. Standard runs use
+        # 1.0 anyway, so this only changes sub-1.0 train-starved runs.
+        sample_ratio = 1.0
         testloader = self.load_data(
             file=self.test_file,
             sample_ratio=sample_ratio,