From 147c8bdbdd29d773a63cb31862b252f31a22b9c6 Mon Sep 17 00:00:00 2001
From: Han Wang <wang_han@iapcm.ac.cn>
Date: Wed, 1 Jul 2026 13:25:24 +0800
Subject: [PATCH] fix(examples): use config-relative data paths in zinc_protein

zinc_se_a_mask.json referenced its training/validation data with
repo-root-relative paths (examples/zinc_protein/...), so the example only
trained when launched from the repository root; from examples/zinc_protein
the paths resolved to examples/zinc_protein/examples/zinc_protein/... and did
not exist. Every other example uses paths relative to its own directory.

Use local train_data_dp_mask/ and val_data_dp_mask/ paths, and extend
test_examples.py to assert that each example's data systems resolve relative
to the example directory (previously only argument-checking was verified, so
unresolvable data paths passed).

Fix #5694
---
 examples/zinc_protein/zinc_se_a_mask.json |  4 ++--
 source/tests/common/test_examples.py      | 23 +++++++++++++++++++++++
 2 files changed, 25 insertions(+), 2 deletions(-)

diff --git a/examples/zinc_protein/zinc_se_a_mask.json b/examples/zinc_protein/zinc_se_a_mask.json
index 8d3c747e08..ae50a8cdc0 100644
--- a/examples/zinc_protein/zinc_se_a_mask.json
+++ b/examples/zinc_protein/zinc_se_a_mask.json
@@ -69,14 +69,14 @@
   "training": {
     "training_data": {
       "systems": [
-        "examples/zinc_protein/train_data_dp_mask/"
+        "train_data_dp_mask/"
       ],
       "batch_size": 2,
       "_comment7": "that's all"
     },
     "validation_data": {
       "systems": [
-        "examples/zinc_protein/val_data_dp_mask/"
+        "val_data_dp_mask/"
       ],
       "batch_size": 2,
       "_comment8": "that's all"
diff --git a/source/tests/common/test_examples.py b/source/tests/common/test_examples.py
index 7c96a7b1ca..e47e6e2837 100644
--- a/source/tests/common/test_examples.py
+++ b/source/tests/common/test_examples.py
@@ -94,3 +94,26 @@ def test_arguments(self) -> None:
                 if multi_task:
                     jdata["model"], _ = preprocess_shared_params(jdata["model"])
                 normalize(jdata, multi_task=multi_task)
+
+    def test_data_paths_exist(self) -> None:
+        """Each example's data ``systems`` must resolve relative to the example's
+        own directory, so the example is runnable from that directory.
+        """
+        for fn in input_files:
+            training = j_loader(str(fn)).get("training", {})
+            for key in ("training_data", "validation_data"):
+                data = training.get(key)
+                if not isinstance(data, dict):
+                    continue
+                systems = data.get("systems")
+                if systems is None:
+                    continue
+                if isinstance(systems, str):
+                    systems = [systems]
+                for system in systems:
+                    with self.subTest(fn=str(fn), key=key, system=system):
+                        self.assertTrue(
+                            (fn.parent / system).exists(),
+                            f"{system!r} in {fn} does not resolve relative to "
+                            f"the example directory {fn.parent}",
+                        )