diff --git a/.github/workflows/test.yml b/.github/workflows/test.yml
index b6a1248..1f5acd7 100644
--- a/.github/workflows/test.yml
+++ b/.github/workflows/test.yml
@@ -34,7 +34,7 @@ jobs:
 
       - name: Install dependencies
         run: |
-          uv pip install cython setuptools --system
+          uv pip install cython 'setuptools<70' --system
 
           python setup.py build_ext --inplace
 
diff --git a/chgnet/trainer/trainer.py b/chgnet/trainer/trainer.py
index 8575b37..3f2ced9 100644
--- a/chgnet/trainer/trainer.py
+++ b/chgnet/trainer/trainer.py
@@ -149,6 +149,7 @@ def __init__(
             )
 
         # Define learning rate scheduler
+        default_decay_frac = 1e-2
         if scheduler in {"MultiStepLR", "multistep"}:
             scheduler_params = kwargs.pop(
                 "scheduler_params",
@@ -164,8 +165,10 @@ def __init__(
             self.scheduler = ExponentialLR(self.optimizer, **scheduler_params)
             self.scheduler_type = "exp"
         elif scheduler in {"CosineAnnealingLR", "CosLR", "Cos", "cos"}:
-            scheduler_params = kwargs.pop("scheduler_params", {"decay_fraction": 1e-2})
-            decay_fraction = scheduler_params.pop("decay_fraction")
+            scheduler_params = kwargs.pop(
+                "scheduler_params", {"decay_fraction": default_decay_frac}
+            )
+            decay_fraction = scheduler_params.pop("decay_fraction", default_decay_frac)
             self.scheduler = CosineAnnealingLR(
                 self.optimizer,
                 T_max=10 * epochs,  # Maximum number of iterations.
@@ -174,9 +177,10 @@ def __init__(
             self.scheduler_type = "cos"
         elif scheduler == "CosRestartLR":
             scheduler_params = kwargs.pop(
-                "scheduler_params", {"decay_fraction": 1e-2, "T_0": 10, "T_mult": 2}
+                "scheduler_params",
+                {"decay_fraction": default_decay_frac, "T_0": 10, "T_mult": 2},
             )
-            decay_fraction = scheduler_params.pop("decay_fraction")
+            decay_fraction = scheduler_params.pop("decay_fraction", default_decay_frac)
             self.scheduler = CosineAnnealingWarmRestarts(
                 self.optimizer,
                 eta_min=decay_fraction * learning_rate,
diff --git a/chgnet/utils/common_utils.py b/chgnet/utils/common_utils.py
index 510594c..4087285 100644
--- a/chgnet/utils/common_utils.py
+++ b/chgnet/utils/common_utils.py
@@ -3,6 +3,7 @@
 import json
 import os
 
+import numpy as np
 import nvidia_smi
 import torch
 from torch import Tensor
@@ -115,8 +116,19 @@ def write_json(dct: dict, filepath: str) -> dict:
     Returns:
         written dictionary
     """
+
+    def handler(obj: object) -> int | object:
+        """Convert numpy int64 to int.
+
+        Fixes TypeError: Object of type int64 is not JSON serializable
+        reported in https://github.com/CederGroupHub/chgnet/issues/168.
+        """
+        if isinstance(obj, np.integer):
+            return int(obj)
+        return obj
+
     with open(filepath, "w") as file:
-        json.dump(dct, file)
+        json.dump(dct, file, default=handler)
 
 
 def mkdir(path: str) -> str:
diff --git a/pyproject.toml b/pyproject.toml
index 470842b..2b20a1d 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -9,7 +9,7 @@ license = { text = "Modified BSD" }
 dependencies = [
     "ase>=3.23.0",
     "cython>=3",
-    "numpy>=1.26",
+    "numpy>=1.26,<2",
     "nvidia-ml-py3>=7.352.0",
     "pymatgen>=2023.10.11",
     "torch>=1.11.0",
@@ -46,7 +46,7 @@ find = { include = ["chgnet*"], exclude = ["tests", "tests*"] }
 "chgnet.pretrained" = ["*", "**/*"]
 
 [build-system]
-requires = ["Cython", "setuptools>=65.0", "wheel"]
+requires = ["Cython", "setuptools>=65,<70", "wheel"]
 build-backend = "setuptools.build_meta"
 
 [tool.ruff]
diff --git a/tests/test_trainer.py b/tests/test_trainer.py
index 9403534..818f193 100644
--- a/tests/test_trainer.py
+++ b/tests/test_trainer.py
@@ -55,15 +55,18 @@ def test_trainer(tmp_path: Path, monkeypatch: pytest.MonkeyPatch) -> None:
         wandb_init_kwargs=dict(anonymous="must"),
         extra_run_config=extra_run_config,
     )
+    trainer.train(
+        train_loader,
+        val_loader,
+        save_dir=tmp_path,
+        save_test_result=tmp_path / "test-preds.json",
+    )
     assert dict(wandb.config).items() >= extra_run_config.items()
-    dir_name = "test_tmp_dir"
-    test_dir = tmp_path / dir_name
-    trainer.train(train_loader, val_loader, save_dir=test_dir)
     for param in chgnet.composition_model.parameters():
         assert param.requires_grad is False
-    assert test_dir.is_dir(), "Training dir was not created"
+    assert tmp_path.is_dir(), "Training dir was not created"
 
-    output_files = [file.name for file in test_dir.iterdir()]
+    output_files = [file.name for file in tmp_path.iterdir()]
     for prefix in ("epoch", "bestE_", "bestF_"):
         n_matches = sum(file.startswith(prefix) for file in output_files)
         assert (
@@ -92,16 +95,14 @@ def test_trainer_composition_model(tmp_path: Path) -> None:
         learning_rate=1e-2,
         epochs=5,
     )
-    dir_name = "test_tmp_dir2"
-    test_dir = tmp_path / dir_name
     initial_weights = chgnet.composition_model.state_dict()["fc.weight"].clone()
     trainer.train(
-        train_loader, val_loader, save_dir=test_dir, train_composition_model=True
+        train_loader, val_loader, save_dir=tmp_path, train_composition_model=True
     )
     for param in chgnet.composition_model.parameters():
         assert param.requires_grad is True
 
-    output_files = list(test_dir.iterdir())
+    output_files = list(tmp_path.iterdir())
     weights_path = next(file for file in output_files if file.name.startswith("epoch"))
     new_chgnet = CHGNet.from_file(weights_path)
     for param in new_chgnet.composition_model.parameters():