[data] fix: make CORD target selection deterministic

Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com>

Author: yaoyu-33

src/megatron/bridge/data/sources/hf_adapters.py

@@ -16,7 +16,6 @@
 
 import io
 import json
-import random
 import re
 from collections.abc import Callable, Iterable, Mapping
 from pathlib import Path
@@ -164,7 +163,9 @@ def _cord_v2_adapter(example: Mapping[str, Any], kwargs: Mapping[str, Any]) -> d
         gt_jsons = ground_truth["gt_parses"]
     else:
         gt_jsons = [ground_truth["gt_parse"]]
-    text = json2token(random.choice(gt_jsons), sort_json_key=True)
+    # Dataset adaptation runs independently on pipeline stages and after RNG
+    # restoration, so target selection must not depend on process-global state.
+    text = json2token(gt_jsons[0], sort_json_key=True)
     prompt = str(kwargs.get("prompt", "Describe this image."))
     return {
         "conversation": [

tests/unit_tests/data/sources/test_hf_adapters.py

@@ -1,6 +1,7 @@
 # Copyright (c) 2026, NVIDIA CORPORATION.  All rights reserved.
 
 import json
+import random
 from types import SimpleNamespace
 
 import pytest
@@ -77,6 +78,20 @@ def test_cord_adapter_supports_plural_and_singular_ground_truth(ground_truth):
     assert adapted[0]["conversation"][1]["role"] == "assistant"
 
 
+def test_cord_adapter_multiple_ground_truths_is_independent_of_global_rng():
+    row = {
+        "image": SimpleNamespace(),
+        "ground_truth": json.dumps({"gt_parses": [{"name": "first"}, {"name": "second"}]}),
+    }
+
+    random.seed(42)
+    first_adaptation = adapt_hf_dataset([row], adapter_name="cord_v2")
+    random.seed(142)
+    second_adaptation = adapt_hf_dataset([row], adapter_name="cord_v2")
+
+    assert first_adaptation == second_adaptation
+
+
 @pytest.mark.parametrize(
     ("adapter_name", "row"),
     [