fix(callback): ensure reset_parameter callback works with sklearn interface

python-package/lightgbm/callback.py

@@ -239,8 +239,9 @@ def __call__(self, env: CallbackEnv) -> None:
                     "Only list and callable values are supported "
                     "as a mapping from boosting round index to new parameter value."
                 )
-            if new_param != env.params.get(key, None):
-                new_parameters[key] = new_param
+            # Always update parameters to ensure compatibility with sklearn interface
+            # The booster's reset_parameter method handles whether the parameter actually changes
+            new_parameters[key] = new_param
         if new_parameters:
             if isinstance(env.model, Booster):
                 env.model.reset_parameter(new_parameters)

python-package/lightgbm/sklearn.py

@@ -975,6 +975,29 @@ def fit(
             else:
                 sample_weight = np.multiply(sample_weight, class_sample_weight)
 
+        # Filter out booster-specific parameters from dataset parameters
+        from .basic import _ConfigAliases
+
+        # Get all booster parameter names to exclude from dataset
+        booster_params = set()
+        # Common booster parameters that should not be in dataset
+        booster_param_names = [
+            'boosting', 'learning_rate', 'num_leaves', 'max_depth', 'min_split_gain',
+            'min_child_weight', 'min_child_samples', 'subsample', 'subsample_freq',
+            'colsample_bytree', 'colsample_bynode', 'colsample_bylevel', 'reg_alpha',
+            'reg_lambda', 'random_state', 'n_estimators', 'num_iterations',
+            'feature_fraction', 'bagging_fraction', 'feature_fraction_bynode',
+            'bagging_freq', 'lambda_l1', 'lambda_l2', 'min_gain_to_split',
+            'min_data_in_leaf', 'min_sum_hessian_in_leaf', 'max_delta_step',
+            'max_bin', 'max_depth', 'min_data_in_bin', 'bin_construct_sample_cnt'
+        ]
+
+        for param_name in booster_param_names:
+            booster_params.update(_ConfigAliases.get_sorted(param_name))
+
+        # Filter dataset parameters
+        dataset_params = {k: v for k, v in params.items() if k not in booster_params}
+
         train_set = Dataset(
             data=_X,
             label=_y,
@@ -983,7 +1006,7 @@ def fit(
             init_score=init_score,
             categorical_feature=categorical_feature,
             feature_name=feature_name,
-            params=params,
+            params=dataset_params,
         )
 
         valid_sets: List[Dataset] = []
@@ -1023,14 +1046,36 @@ def fit(
                         name="eval_group",
                         i=i,
                     )
+                    # Filter out booster-specific parameters from dataset parameters
+                    from .basic import _ConfigAliases
+
+                    # Get all booster parameter names to exclude from dataset
+                    booster_params = set()
+                    booster_param_names = [
+                        'boosting', 'learning_rate', 'num_leaves', 'max_depth', 'min_split_gain',
+                        'min_child_weight', 'min_child_samples', 'subsample', 'subsample_freq',
+                        'colsample_bytree', 'colsample_bynode', 'colsample_bylevel', 'reg_alpha',
+                        'reg_lambda', 'random_state', 'n_estimators', 'num_iterations',
+                        'feature_fraction', 'bagging_fraction', 'feature_fraction_bynode',
+                        'bagging_freq', 'lambda_l1', 'lambda_l2', 'min_gain_to_split',
+                        'min_data_in_leaf', 'min_sum_hessian_in_leaf', 'max_delta_step',
+                        'max_bin', 'max_depth', 'min_data_in_bin', 'bin_construct_sample_cnt'
+                    ]
+
+                    for param_name in booster_param_names:
+                        booster_params.update(_ConfigAliases.get_sorted(param_name))
+
+                    # Filter dataset parameters
+                    valid_dataset_params = {k: v for k, v in params.items() if k not in booster_params}
+
                     valid_set = Dataset(
                         data=valid_data[0],
                         label=valid_data[1],
                         weight=valid_weight,
                         group=valid_group,
                         init_score=valid_init_score,
                         categorical_feature="auto",
-                        params=params,
+                        params=valid_dataset_params,
                     )
 
                 valid_sets.append(valid_set)

tests/python_package_test/test_callback.py

@@ -63,3 +63,29 @@ def test_reset_parameter_callback_is_picklable(serializer):
     assert callback_from_disk.before_iteration is True
     assert callback.kwargs == callback_from_disk.kwargs
     assert callback.kwargs == params
+
+
+def test_reset_parameter_callback_with_sklearn():
+    """Test that reset_parameter callback works with LGBMClassifier."""
+    import numpy as np
+    import lightgbm as lgb
+    from lightgbm import LGBMClassifier
+    from sklearn.datasets import make_classification
+
+    X, y = make_classification(n_samples=1000, n_features=10, random_state=42)
+
+    model = LGBMClassifier(
+        n_estimators=10,
+        colsample_bytree=0.9,  # Start high
+        callbacks=[lgb.reset_parameter(colsample_bytree=[0.3, 0.8, 0.3, 0.8, 0.3, 0.8, 0.3, 0.8, 0.3, 0.8])],
+        verbose=-1
+    )
+    model.fit(X, y)
+
+    trees_df = model.booster_.trees_to_dataframe()
+    unique_feature_counts = trees_df.groupby('tree_index')['split_feature'].nunique()
+
+    assert unique_feature_counts.nunique() > 1, (
+        f"reset_parameter callback did not work with LGBMClassifier. "
+        f"All trees used the same number of features. Counts: {unique_feature_counts.unique()}"
+    )