diff --git a/MIGRATION.md b/MIGRATION.md
index 6877b331c72..ba2c8e2338a 100644
--- a/MIGRATION.md
+++ b/MIGRATION.md
@@ -15,6 +15,26 @@ This guide covers the breaking changes introduced in TRL v1 and how to update yo
 | --- | --- | --- | --- | --- |
 | `SFTConfig` | `packing` | `"bfd-requeue"` | `"bfd_split"` | Replace `packing="bfd-requeue"` with `packing="bfd_split"`. The old value will still be accepted for a few versions but will be removed in a future release. |
 
+## Removed automatic `None` stripping from trainer preprocessing
+
+TRL trainers (SFT, DPO, Reward) no longer automatically strip `None` values from dataset examples during preprocessing. Previously, each trainer applied `remove_none_values` via `dataset.with_transform` to work around tabular backends (Arrow/Parquet) inserting `None` for missing keys in nested structures.
+
+This affects datasets that contain `None` values because they were:
+
+- Created before `datasets` v4.7.0, which introduced the Json dtype that preserves nested structures without inserting `None`.
+- Created with `datasets` v4.7.0 or later, but saved without using the Json feature.
+
+**Action needed:** If your dataset falls into one of the above categories and contains `None` values in nested columns, apply the fix manually before training:
+
+```python
+from trl.trainer.utils import remove_none_values
+
+dataset = dataset.with_transform(remove_none_values)
+trainer = SFTTrainer(..., train_dataset=dataset)
+```
+
+Datasets created or re-saved with `datasets` v4.7.0+ using the Json dtype are unaffected.
+
 ## Migrating from an earlier version
 
 Depending on which version you're migrating from, refer to the [release notes](https://github.com/huggingface/trl/releases) for v0.29 and earlier for version-specific changes.
diff --git a/pyproject.toml b/pyproject.toml
index ac9f4cc991f..f4831765478 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -30,7 +30,7 @@ classifiers = [
 requires-python = ">=3.10"
 dependencies = [
     "accelerate>=1.4.0",
-    "datasets>=3.0.0",
+    "datasets>=4.7.0",  # Support Json type and on_mixed_types="use_json"
     "packaging>20.0",
     "transformers>=4.56.2",
 ]
diff --git a/scripts/generate_harmony_dataset.py b/scripts/generate_harmony_dataset.py
index 670c586a4bb..2257be8b7e6 100644
--- a/scripts/generate_harmony_dataset.py
+++ b/scripts/generate_harmony_dataset.py
@@ -91,7 +91,7 @@ def main(test_size, push_to_hub, repo_id):
             {"reasoning_effort": "high", "model_identity": "You are Tiny ChatGPT, a tiny language model."},
             {"reasoning_effort": "low", "model_identity": "You are Tiny ChatGPT, a tiny language model."},
         ]
-    })
+    }, on_mixed_types="use_json")
     language_modeling_dataset = language_modeling_dataset.train_test_split(test_size=test_size, shuffle=False)
     if push_to_hub:
         language_modeling_dataset.push_to_hub(repo_id, config_name="language_modeling")
diff --git a/scripts/generate_toolcall_dataset.py b/scripts/generate_toolcall_dataset.py
index d3c9caa88d5..84b56b88ab9 100644
--- a/scripts/generate_toolcall_dataset.py
+++ b/scripts/generate_toolcall_dataset.py
@@ -12,7 +12,6 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-import json
 from dataclasses import dataclass, field
 
 from datasets import Dataset
@@ -213,16 +212,16 @@ def get_wind_conditions(city: str, unit: str) -> tuple[int, str]:
             ]
         ],
         "tools": [
-            json.dumps([start_timer, create_reminder]),
-            json.dumps([get_current_time]),
-            json.dumps([get_air_quality_index, get_weather_forecast, get_wind_conditions]),
-            json.dumps([play_music, control_light]),
-            json.dumps([get_weather_forecast, get_wind_conditions]),
-            json.dumps([control_light]),
-            json.dumps([start_timer, create_reminder]),
-            json.dumps([get_weather_forecast, get_wind_conditions]),
+            [start_timer, create_reminder],
+            [get_current_time],
+            [get_air_quality_index, get_weather_forecast, get_wind_conditions],
+            [play_music, control_light],
+            [get_weather_forecast, get_wind_conditions],
+            [control_light],
+            [start_timer, create_reminder],
+            [get_weather_forecast, get_wind_conditions],
         ]
-    })
+    }, on_mixed_types="use_json")
     language_modeling_dataset = language_modeling_dataset.train_test_split(test_size=test_size, shuffle=False)
     if push_to_hub:
         language_modeling_dataset.push_to_hub(repo_id, config_name="language_modeling")
@@ -319,16 +318,16 @@ def get_wind_conditions(city: str, unit: str) -> tuple[int, str]:
             ],
         ],
         "tools": [
-            json.dumps([start_timer]),
-            json.dumps([get_current_time]),
-            json.dumps([get_air_quality_index]),
-            json.dumps([play_music]),
-            json.dumps([get_weather_forecast]),
-            json.dumps([control_light]),
-            json.dumps([create_reminder]),
-            json.dumps([get_wind_conditions]),
+            [start_timer],
+            [get_current_time],
+            [get_air_quality_index],
+            [play_music],
+            [get_weather_forecast],
+            [control_light],
+            [create_reminder],
+            [get_wind_conditions],
         ],
-    })
+    }, on_mixed_types="use_json")
     preference_dataset = preference_dataset.train_test_split(test_size=test_size, shuffle=False)
     if push_to_hub:
         preference_dataset.push_to_hub(repo_id, config_name="preference")
diff --git a/trl/trainer/dpo_trainer.py b/trl/trainer/dpo_trainer.py
index d175d09b5ec..29f9102a0e3 100644
--- a/trl/trainer/dpo_trainer.py
+++ b/trl/trainer/dpo_trainer.py
@@ -59,7 +59,6 @@
     get_config_model_id,
     hash_module,
     pad,
-    remove_none_values,
     selective_log_softmax,
     use_adapter,
 )
@@ -863,11 +862,6 @@ def _prepare_dataset(
         args: DPOConfig,
         dataset_name: str,
     ) -> Dataset | IterableDataset:
-        # Tabular backends like Arrow/Parquet insert `None` for mismatched keys in nested structures. Clean them from
-        # sampled data.
-        if isinstance(dataset, Dataset):  # IterableDataset does not support `with_transform`
-            dataset = dataset.with_transform(remove_none_values)
-
         # Build the kwargs for the `map` function
         map_kwargs = {}
         if isinstance(dataset, Dataset):  # IterableDataset does not support num_proc
diff --git a/trl/trainer/reward_trainer.py b/trl/trainer/reward_trainer.py
index cd079499b51..5f19c1eb907 100644
--- a/trl/trainer/reward_trainer.py
+++ b/trl/trainer/reward_trainer.py
@@ -52,7 +52,7 @@
 from ..models import get_act_offloading_ctx_manager
 from .base_trainer import _BaseTrainer
 from .reward_config import RewardConfig
-from .utils import create_model_from_path, disable_dropout_in_model, get_config_model_id, pad, remove_none_values
+from .utils import create_model_from_path, disable_dropout_in_model, get_config_model_id, pad
 
 
 if is_peft_available():
@@ -541,11 +541,6 @@ def _prepare_dataset(
         args: RewardConfig,
         dataset_name: str,
     ) -> Dataset | IterableDataset:
-        # Tabular backends like Arrow/Parquet insert `None` for mismatched keys in nested structures. Clean them from
-        # sampled data.
-        if isinstance(dataset, Dataset):  # IterableDataset does not support `with_transform`
-            dataset = dataset.with_transform(remove_none_values)
-
         # If the dataset is already preprocessed (tokenized), skip the processing steps.
         column_names = get_dataset_column_names(dataset)
         is_processed = "chosen_ids" in column_names and "rejected_ids" in column_names
diff --git a/trl/trainer/sft_trainer.py b/trl/trainer/sft_trainer.py
index 7207b98349a..59fb77b5707 100644
--- a/trl/trainer/sft_trainer.py
+++ b/trl/trainer/sft_trainer.py
@@ -61,7 +61,6 @@
     flush_left,
     get_config_model_id,
     pad,
-    remove_none_values,
     selective_log_softmax,
 )
 
@@ -1054,11 +1053,6 @@ def _prepare_dataset(
         formatting_func: Callable[[dict], str] | None,
         dataset_name: str,
     ) -> Dataset | IterableDataset:
-        # Tabular backends like Arrow/Parquet insert `None` for mismatched keys in nested structures. Clean them from
-        # sampled data.
-        if isinstance(dataset, Dataset):  # IterableDataset does not support `with_transform`
-            dataset = dataset.with_transform(remove_none_values)
-
         # If the dataset is already preprocessed (tokenized), skip the processing steps.
         column_names = get_dataset_column_names(dataset)
         is_processed = "input_ids" in column_names
diff --git a/trl/trainer/utils.py b/trl/trainer/utils.py
index 1db504acebe..7e22e241c1a 100644
--- a/trl/trainer/utils.py
+++ b/trl/trainer/utils.py
@@ -972,6 +972,12 @@ def unsplit_pixel_values_by_grid(batch: dict[str, torch.Tensor | list[torch.Tens
 TListOrMapping = TypeVar("TListOrMapping", list, Mapping)
 
 
+# This function is intentionally not used internally. It is provided as a utility for users whose datasets contain
+# `None` values inserted by tabular backends (e.g., Arrow/Parquet) for missing keys in nested structures. This
+# situation arises when loading datasets created before `datasets` v4.7.0 (which introduced the Json dtype), or when
+# datasets created after that version were saved without using the Json feature. In both cases, users can apply this
+# function via `dataset = dataset.with_transform(remove_none_values)` before training to strip the spurious `None`
+# values. See the migration guide for more details.
 def remove_none_values(example: TListOrMapping) -> TListOrMapping:
     """
     Recursively removes entries with `None` values from a nested structure (list or dictionary).
@@ -980,7 +986,10 @@ def remove_none_values(example: TListOrMapping) -> TListOrMapping:
         example (`list` or `Mapping`):
             Input nested structure (list or dictionary) from which to remove `None`.
 
-    Example:
+    Examples:
+    ```python
+    >>> dataset = dataset.with_transform(remove_none_values)
+    ```
     ```python
     >>> [
     ...     {