diff --git a/python-package/xgboost/data.py b/python-package/xgboost/data.py
index d49ff5e43899..9ce2b6b17e1c 100644
--- a/python-package/xgboost/data.py
+++ b/python-package/xgboost/data.py
@@ -995,11 +995,7 @@ def _from_uri(
     _warn_unused_missing(data, missing)
     handle = ctypes.c_void_p()
     data = os.fspath(os.path.expanduser(data))
-    args = {
-        "uri": str(data),
-        "data_split_mode": int(data_split_mode),
-    }
-    config = bytes(json.dumps(args), "utf-8")
+    config = make_jcargs(uri=str(data), data_split_mode=int(data_split_mode))
     _check_call(_LIB.XGDMatrixCreateFromURI(config, ctypes.byref(handle)))
     return handle, feature_names, feature_types
 
diff --git a/python-package/xgboost/testing/__init__.py b/python-package/xgboost/testing/__init__.py
index 53779403917b..29feac569b61 100644
--- a/python-package/xgboost/testing/__init__.py
+++ b/python-package/xgboost/testing/__init__.py
@@ -37,20 +37,19 @@
 import xgboost as xgb
 from xgboost import RabitTracker
 from xgboost.core import ArrayLike
-from xgboost.data import is_pd_cat_dtype
 from xgboost.sklearn import SklObjective
-from xgboost.testing.data import (
+
+from .._typing import PathLike
+from .data import (
     get_california_housing,
     get_cancer,
     get_digits,
     get_sparse,
     make_batches,
+    make_categorical,
     make_sparse_regression,
-    memory,
 )
 
-from .._typing import PathLike
-
 hypothesis = pytest.importorskip("hypothesis")
 
 # pylint:disable=wrong-import-position,wrong-import-order
@@ -377,81 +376,6 @@ def __repr__(self) -> str:
         return self.name
 
 
-# pylint: disable=too-many-arguments,too-many-locals
-@memory.cache
-def make_categorical(
-    n_samples: int,
-    n_features: int,
-    n_categories: int,
-    *,
-    onehot: bool,
-    sparsity: float = 0.0,
-    cat_ratio: float = 1.0,
-    shuffle: bool = False,
-    random_state: int = 1994,
-) -> Tuple[ArrayLike, np.ndarray]:
-    """Generate categorical features for test.
-
-    Parameters
-    ----------
-    n_categories:
-        Number of categories for categorical features.
-    onehot:
-        Should we apply one-hot encoding to the data?
-    sparsity:
-        The ratio of the amount of missing values over the number of all entries.
-    cat_ratio:
-        The ratio of features that are categorical.
-    shuffle:
-        Whether we should shuffle the columns.
-
-    Returns
-    -------
-    X, y
-    """
-    import pandas as pd
-
-    rng = np.random.RandomState(random_state)
-
-    pd_dict = {}
-    for i in range(n_features + 1):
-        c = rng.randint(low=0, high=n_categories, size=n_samples)
-        pd_dict[str(i)] = pd.Series(c, dtype=np.int64)
-
-    df = pd.DataFrame(pd_dict)
-    label = df.iloc[:, 0]
-    df = df.iloc[:, 1:]
-    for i in range(0, n_features):
-        label += df.iloc[:, i]
-    label += 1
-
-    categories = np.arange(0, n_categories)
-    for col in df.columns:
-        if rng.binomial(1, cat_ratio, size=1)[0] == 1:
-            df[col] = df[col].astype("category")
-            df[col] = df[col].cat.set_categories(categories)
-
-    if sparsity > 0.0:
-        for i in range(n_features):
-            index = rng.randint(
-                low=0, high=n_samples - 1, size=int(n_samples * sparsity)
-            )
-            df.iloc[index, i] = np.nan
-            if is_pd_cat_dtype(df.dtypes.iloc[i]):
-                assert n_categories == np.unique(df.dtypes.iloc[i].categories).size
-
-    assert df.shape[1] == n_features
-    if onehot:
-        df = pd.get_dummies(df)
-
-    if shuffle:
-        columns = list(df.columns)
-        rng.shuffle(columns)
-        df = df[columns]
-
-    return df, label
-
-
 def make_ltr(
     n_samples: int,
     n_features: int,
diff --git a/python-package/xgboost/testing/data.py b/python-package/xgboost/testing/data.py
index 34f55c077a85..4ad5915aad88 100644
--- a/python-package/xgboost/testing/data.py
+++ b/python-package/xgboost/testing/data.py
@@ -1,7 +1,8 @@
-# pylint: disable=invalid-name
+# pylint: disable=invalid-name, too-many-lines
 """Utilities for data generation."""
 import multiprocessing
 import os
+import string
 import zipfile
 from concurrent.futures import ThreadPoolExecutor
 from dataclasses import dataclass
@@ -14,6 +15,7 @@
     List,
     NamedTuple,
     Optional,
+    Set,
     Tuple,
     Type,
     Union,
@@ -26,8 +28,10 @@
 from numpy.random import Generator as RNG
 from scipy import sparse
 
-import xgboost
-from xgboost.data import pandas_pyarrow_mapper
+from ..core import DMatrix, QuantileDMatrix
+from ..data import is_pd_cat_dtype, pandas_pyarrow_mapper
+from ..sklearn import ArrayLike, XGBRanker
+from ..training import train as train_fn
 
 if TYPE_CHECKING:
     from ..compat import DataFrame as DataFrameT
@@ -42,7 +46,7 @@ def np_dtypes(
     n_samples: int, n_features: int
 ) -> Generator[Tuple[np.ndarray, np.ndarray], None, None]:
     """Enumerate all supported dtypes from numpy."""
-    import pandas as pd
+    pd = pytest.importorskip("pandas")
 
     rng = np.random.RandomState(1994)
     # Integer and float.
@@ -99,7 +103,7 @@ def np_dtypes(
 
 def pd_dtypes() -> Generator:
     """Enumerate all supported pandas extension types."""
-    import pandas as pd
+    pd = pytest.importorskip("pandas")
 
     # Integer
     dtypes = [
@@ -162,8 +166,8 @@ def pd_dtypes() -> Generator:
 
 def pd_arrow_dtypes() -> Generator:
     """Pandas DataFrame with pyarrow backed type."""
-    import pandas as pd
-    import pyarrow as pa
+    pd = pytest.importorskip("pandas")
+    pa = pytest.importorskip("pyarrow")
 
     # Integer
     dtypes = pandas_pyarrow_mapper
@@ -225,10 +229,10 @@ def check_inf(rng: RNG) -> None:
     X[5, 2] = np.inf
 
     with pytest.raises(ValueError, match="Input data contains `inf`"):
-        xgboost.QuantileDMatrix(X, y)
+        QuantileDMatrix(X, y)
 
     with pytest.raises(ValueError, match="Input data contains `inf`"):
-        xgboost.DMatrix(X, y)
+        DMatrix(X, y)
 
 
 @memory.cache
@@ -288,8 +292,10 @@ def get_ames_housing() -> Tuple[DataFrameT, np.ndarray]:
     Number of categorical features: 10
     Number of numerical features: 10
     """
-    pytest.importorskip("pandas")
-    import pandas as pd
+    if TYPE_CHECKING:
+        import pandas as pd
+    else:
+        pd = pytest.importorskip("pandas")
 
     rng = np.random.default_rng(1994)
     n_samples = 1460
@@ -664,7 +670,7 @@ def init_rank_score(
     y_train = y_train[sorted_idx]
     qid_train = qid_train[sorted_idx]
 
-    ltr = xgboost.XGBRanker(objective="rank:ndcg", tree_method="hist")
+    ltr = XGBRanker(objective="rank:ndcg", tree_method="hist")
     ltr.fit(X_train, y_train, qid=qid_train)
 
     # Use the original order of the data.
@@ -799,9 +805,7 @@ def sort_ltr_samples(
     return data
 
 
-def run_base_margin_info(
-    DType: Callable, DMatrixT: Type[xgboost.DMatrix], device: str
-) -> None:
+def run_base_margin_info(DType: Callable, DMatrixT: Type[DMatrix], device: str) -> None:
     """Run tests for base margin."""
     rng = np.random.default_rng()
     X = DType(rng.normal(0, 1.0, size=100).astype(np.float32).reshape(50, 2))
@@ -814,7 +818,7 @@ def run_base_margin_info(
     Xy = DMatrixT(X, y, base_margin=base_margin)
     # Error at train, caused by check in predictor.
     with pytest.raises(ValueError, match=r".*base_margin.*"):
-        xgboost.train({"tree_method": "hist", "device": device}, Xy)
+        train_fn({"tree_method": "hist", "device": device}, Xy)
 
     if not hasattr(X, "iloc"):
         # column major matrix
@@ -932,3 +936,102 @@ def random_csc(t_id: int) -> sparse.csc_matrix:
         return arr, y
 
     return csr, y
+
+
+def unique_random_strings(n_strings: int, seed: int) -> List[str]:
+    """Generate n unique strings."""
+    name_len = 8  # hardcoded, should be more than enough
+    unique_strings: Set[str] = set()
+    rng = np.random.default_rng(seed)
+
+    while len(unique_strings) < n_strings:
+        random_str = "".join(
+            rng.choice(list(string.ascii_letters), size=name_len, replace=True)
+        )
+        unique_strings.add(random_str)
+
+    return list(unique_strings)
+
+
+# pylint: disable=too-many-arguments,too-many-locals,too-many-branches
+def make_categorical(
+    n_samples: int,
+    n_features: int,
+    n_categories: int,
+    *,
+    onehot: bool,
+    sparsity: float = 0.0,
+    cat_ratio: float = 1.0,
+    shuffle: bool = False,
+    random_state: int = 1994,
+    cat_dtype: np.typing.DTypeLike = np.int64,
+) -> Tuple[ArrayLike, np.ndarray]:
+    """Generate categorical features for test.
+
+    Parameters
+    ----------
+    n_categories:
+        Number of categories for categorical features.
+    onehot:
+        Should we apply one-hot encoding to the data?
+    sparsity:
+        The ratio of the amount of missing values over the number of all entries.
+    cat_ratio:
+        The ratio of features that are categorical.
+    shuffle:
+        Whether we should shuffle the columns.
+    cat_dtype :
+        The dtype for categorical features, might be string or numeric.
+
+    Returns
+    -------
+    X, y
+    """
+    pd = pytest.importorskip("pandas")
+
+    rng = np.random.RandomState(random_state)
+
+    df = pd.DataFrame()
+    for i in range(n_features):
+        choice = rng.binomial(1, cat_ratio, size=1)[0]
+        if choice == 1:
+            if np.issubdtype(cat_dtype, np.str_):
+                categories = np.array(unique_random_strings(n_categories, i))
+                c = rng.choice(categories, size=n_samples, replace=True)
+            else:
+                categories = np.arange(0, n_categories)
+                c = rng.randint(low=0, high=n_categories, size=n_samples)
+
+            df[str(i)] = pd.Series(c, dtype="category")
+            df[str(i)] = df[str(i)].cat.set_categories(categories)
+        else:
+            num = rng.randint(low=0, high=n_categories, size=n_samples)
+            df[str(i)] = pd.Series(num, dtype=num.dtype)
+
+    label = np.zeros(shape=(n_samples,))
+    for col in df.columns:
+        if isinstance(df[col].dtype, pd.CategoricalDtype):
+            label += df[col].cat.codes
+        else:
+            label += df[col]
+    label += 1
+
+    if sparsity > 0.0:
+        for i in range(n_features):
+            index = rng.randint(
+                low=0, high=n_samples - 1, size=int(n_samples * sparsity)
+            )
+            df.iloc[index, i] = np.nan
+            if is_pd_cat_dtype(df.dtypes.iloc[i]):
+                assert n_categories == np.unique(df.dtypes.iloc[i].categories).size
+
+    assert df.shape[1] == n_features
+    if onehot:
+        df = pd.get_dummies(df)
+
+    if shuffle:
+        columns = list(df.columns)
+        rng.shuffle(columns)
+        df = df[columns]
+
+    return df, label
diff --git a/python-package/xgboost/testing/quantile_dmatrix.py b/python-package/xgboost/testing/quantile_dmatrix.py
index b06cb550198f..f5e861c50098 100644
--- a/python-package/xgboost/testing/quantile_dmatrix.py
+++ b/python-package/xgboost/testing/quantile_dmatrix.py
@@ -1,11 +1,12 @@
 """QuantileDMatrix related tests."""
 
 import numpy as np
+import pytest
 from sklearn.model_selection import train_test_split
 
 import xgboost as xgb
 
-from .data import make_batches
+from .data import make_batches, make_categorical
 
 
 def check_ref_quantile_cut(device: str) -> None:
@@ -33,3 +34,35 @@ def check_ref_quantile_cut(device: str) -> None:
     Xy_valid = xgb.QuantileDMatrix(X_valid, y_valid)
     cut_valid = Xy_valid.get_quantile_cut()
     assert not np.allclose(cut_train[1], cut_valid[1])
+
+
+def check_categorical_strings(device: str) -> None:
+    """Check string inputs."""
+    if device == "cpu":
+        pd = pytest.importorskip("pandas")
+    else:
+        pd = pytest.importorskip("cudf")
+
+    n_categories = 32
+    X, y = make_categorical(
+        1024,
+        8,
+        n_categories,
+        onehot=False,
+        cat_dtype=np.str_,
+        cat_ratio=0.5,
+        shuffle=True,
+    )
+    X = pd.DataFrame(X)
+
+    Xy = xgb.QuantileDMatrix(X, y, enable_categorical=True)
+    assert Xy.num_col() == 8
+    cuts = Xy.get_quantile_cut()
+    indptr = cuts[0]
+    values = cuts[1]
+    for i in range(1, len(indptr)):
+        f_idx = i - 1
+        if isinstance(X[X.columns[f_idx]].dtype, pd.CategoricalDtype):
+            beg, end = indptr[f_idx], indptr[i]
+            col = values[beg:end]
+            np.testing.assert_allclose(col, np.arange(0, n_categories))
diff --git a/tests/python-gpu/test_device_quantile_dmatrix.py b/tests/python-gpu/test_device_quantile_dmatrix.py
index 2f2e6545bf2d..5e3bd79625e5 100644
--- a/tests/python-gpu/test_device_quantile_dmatrix.py
+++ b/tests/python-gpu/test_device_quantile_dmatrix.py
@@ -8,7 +8,10 @@
 from xgboost import testing as tm
 from xgboost.testing.data import check_inf
 from xgboost.testing.data_iter import run_mixed_sparsity
-from xgboost.testing.quantile_dmatrix import check_ref_quantile_cut
+from xgboost.testing.quantile_dmatrix import (
+    check_categorical_strings,
+    check_ref_quantile_cut,
+)
 
 sys.path.append("tests/python")
 import test_quantile_dmatrix as tqd
@@ -33,6 +36,9 @@ def test_dmatrix_feature_weights(self) -> None:
             feature_weights.astype(np.float32),
         )
 
+    def test_categorical_strings(self) -> None:
+        check_categorical_strings("cuda")
+
     @pytest.mark.skipif(**tm.no_cupy())
     def test_dmatrix_cupy_init(self) -> None:
         import cupy as cp
diff --git a/tests/python-gpu/test_from_cudf.py b/tests/python-gpu/test_from_cudf.py
index 37826f35cc34..0d138f14b29a 100644
--- a/tests/python-gpu/test_from_cudf.py
+++ b/tests/python-gpu/test_from_cudf.py
@@ -210,8 +210,8 @@ def test_cudf_categorical(self) -> None:
         assert all(t == "c" for t in Xy.feature_types)
 
         # mixed dtypes
-        X["1"] = X["1"].astype(np.int64)
-        X["3"] = X["3"].astype(np.int64)
+        X["0"] = X["0"].astype(np.int64)
+        X["2"] = X["2"].astype(np.int64)
         df, cat_codes, _, _ = xgb.data._transform_cudf_df(
             X, None, None, enable_categorical=True
         )
diff --git a/tests/python/test_data_iterator.py b/tests/python/test_data_iterator.py
index 545b849b4bdb..6a980f967a97 100644
--- a/tests/python/test_data_iterator.py
+++ b/tests/python/test_data_iterator.py
@@ -259,13 +259,13 @@ def test_cat_check() -> None:
     batches = []
 
     for i in range(n_batches):
-        X, y = tm.make_categorical(
+        X_df, y_arr = tm.make_categorical(
             n_samples=n_samples_per_batch,
             n_features=n_features,
             n_categories=3,
             onehot=False,
         )
-        batches.append((X, y))
+        batches.append((X_df, y_arr))
 
     X, y = list(zip(*batches))
     it = tm.IteratorForTest(X, y, None, cache=None, on_host=False)
diff --git a/tests/python/test_quantile_dmatrix.py b/tests/python/test_quantile_dmatrix.py
index e1152370732b..e64212265212 100644
--- a/tests/python/test_quantile_dmatrix.py
+++ b/tests/python/test_quantile_dmatrix.py
@@ -17,7 +17,10 @@
 )
 from xgboost.testing.data import check_inf, np_dtypes
 from xgboost.testing.data_iter import run_mixed_sparsity
-from xgboost.testing.quantile_dmatrix import check_ref_quantile_cut
+from xgboost.testing.quantile_dmatrix import (
+    check_categorical_strings,
+    check_ref_quantile_cut,
+)
 
 
 class TestQuantileDMatrix:
@@ -57,6 +60,9 @@ def test_basic(self) -> None:
         r = np.arange(1.0, n_samples)
         np.testing.assert_allclose(Xy.get_data().toarray()[1:, 0], r)
 
+    def test_categorical_strings(self) -> None:
+        check_categorical_strings("cpu")
+
     def test_error(self):
         from sklearn.model_selection import train_test_split