[c++] raise a clear error for unknown metric types

.ci/test-r-package.sh

@@ -49,6 +49,7 @@ if [[ $OS_NAME == "linux" ]]; then
         --no-install-recommends \
         -y \
             devscripts \
+            libuv1-dev \
             r-base-core=${R_LINUX_VERSION} \
             r-base-dev=${R_LINUX_VERSION} \
             texinfo \

.github/workflows/r_package.yml

@@ -273,15 +273,16 @@ jobs:
         # references:
         #   * CRAN "additional checks": https://cran.r-project.org/web/checks/check_issue_kinds.html
         #   * images: https://r-hub.github.io/containers/containers.html
+        #
+        # Generally, only images that are still supported, listed at
+        # https://r-hub.github.io/containers/#available-containers, should be used.
         image:
-          - clang16
-          - clang17
-          - clang18
-          - clang19
-          - clang20
-          - gcc14
-          - intel
+          - clang21
+          - clang22
+          - gcc16
           - rchk
+          - ubuntu-clang
+          - ubuntu-gcc12
     runs-on: ubuntu-latest
     container: ghcr.io/r-hub/containers/${{ matrix.image }}:latest
     steps:
@@ -302,6 +303,7 @@ jobs:
             apt-get install --no-install-recommends -y \
                 cmake \
                 devscripts \
+                libcurl4-openssl-dev \
                 libuv1-dev \
                 texinfo \
                 texlive-latex-extra \
@@ -315,10 +317,15 @@ jobs:
             yum install -y \
                 cmake \
                 devscripts \
+                libcurl-devel \
                 libuv-devel \
                 qpdf \
                 texinfo \
                 texinfo-tex \
+                texlive-amsmath \
+                texlive-amsfonts \
+                texlive-collection-fontsrecommended \
+                texlive-collection-latexrecommended \
                 texlive-latex \
                 tidy
           fi
@@ -348,8 +355,16 @@ jobs:
             fi
           fi
 
-          # 'testthat' is not needed by 'rchk', so avoid installing it until here
-          Rscript -e "install.packages('testthat', repos = 'https://cran.rstudio.com', Ncpus = parallel::detectCores())"
+          # why these packages?
+          #
+          #   - 'curl' is not a dependency of {lightgbm}, but it's needed for 'R CMD check' URL checks
+          #   - 'testthat' is not needed by 'rchk', so avoid installing it until here
+          #
+          Rscript -e "
+            install.packages(c('curl', 'testthat'),
+            repos = 'https://cran.rstudio.com',
+            Ncpus = parallel::detectCores())
+          "
 
           if [[ "${{ matrix.image }}" =~ "clang" ]]; then
             # allowing the following NOTEs (produced by default in the clang images):

R-package/tests/testthat/test_basic.R

@@ -765,6 +765,52 @@ test_that("lgb.train() raises an informative error for unrecognized objectives",
   }, regexp = "Unknown objective type name: not_a_real_objective")
 })
 
+test_that("lgb.train() raises an informative error for unrecognized metrics", {
+  dtrain <- lgb.Dataset(
+    data = train$data
+    , label = train$label
+    , params = list(num_threads = .LGB_MAX_THREADS)
+  )
+  expect_error({
+    capture.output({
+      bst <- lgb.train(
+        data = dtrain
+        , valids = list(valid = dtrain)
+        , params = list(
+          objective = "binary"
+          , metric = "nonsense"
+          , verbosity = .LGB_VERBOSITY
+          , num_threads = .LGB_MAX_THREADS
+          , num_iterations = 2L
+        )
+      )
+    }, type = "message")
+  }, regexp = "Unknown metric 'nonsense'")
+})
+
+test_that("lgb.cv() raises an informative error for unrecognized metrics", {
+  dtrain <- lgb.Dataset(
+    data = train$data
+    , label = train$label
+    , params = list(num_threads = .LGB_MAX_THREADS)
+  )
+  expect_error({
+    capture.output({
+      bst <- lgb.cv(
+        data = dtrain
+        , params = list(
+          objective = "binary"
+          , metric = "nonsense"
+          , verbosity = .LGB_VERBOSITY
+          , num_threads = .LGB_MAX_THREADS
+        )
+        , nfold = 2L
+        , nrounds = 2L
+      )
+    }, type = "message")
+  }, regexp = "Unknown metric 'nonsense'")
+})
+
 test_that("lgb.train() respects parameter aliases for objective", {
   nrounds <- 3L
   dtrain <- lgb.Dataset(

python-package/lightgbm/basic.py

@@ -3115,7 +3115,8 @@ def set_weight(
         # Check if the weight contains values other than one
         if weight is not None:
             if _is_pyarrow_array(weight):
-                if pa_compute.all(pa_compute.equal(weight, 1)).as_py():
+                # TODO: remove 'type: ignore[attr-defined]' when https://github.com/apache/arrow/issues/49831 is resolved.
+                if pa_compute.all(pa_compute.equal(weight, 1)).as_py():  # type: ignore[attr-defined]
                     weight = None
             elif np.all(weight == 1):
                 weight = None

src/metric/metric.cpp

@@ -136,6 +136,12 @@ Metric* Metric::CreateMetric(const std::string& type, const Config& config) {
   #ifdef USE_CUDA
   }
   #endif  // USE_CUDA
+  // "custom" is a sentinel value produced by ParseMetricAlias for "none" / "na" / "null" / "custom";
+  // it means the user deliberately requested no built-in metric, so returning nullptr is intentional.
+  // Any other unrecognised name is a user mistake and deserves a clear error.
+  if (type != std::string("custom")) {
+    Log::Fatal("Unknown metric '%s'. To avoid enabling LightGBM built-in metrics, pass metric='none'. Otherwise, ensure 'metric' contains only supported values.", type.c_str());
+  }
   return nullptr;
 }
 

tests/python_package_test/test_engine.py

@@ -4811,6 +4811,42 @@ def test_train_raises_informative_error_for_params_of_wrong_type():
         lgb.train(params, dtrain)
 
 
+def test_train_raises_informative_error_for_unknown_metric():
+    X, y = load_breast_cancer(return_X_y=True)
+    dtrain = lgb.Dataset(X, label=y)
+    params = {"objective": "binary", "metric": "nonsense", "verbose": -1}
+    with pytest.raises(lgb.basic.LightGBMError, match="Unknown metric 'nonsense'"):
+        lgb.train(params, dtrain, num_boost_round=1, valid_sets=[dtrain])
+
+
+@pytest.mark.parametrize(
+    "metric",
+    [
+        "none",
+        ["none", "null"],
+        ["None"],
+        ["NULL"],
+    ],
+)
+def test_train_sentinel_metric_disables_builtin_metrics(metric):
+    X, y = load_breast_cancer(return_X_y=True)
+    dtrain = lgb.Dataset(X, label=y)
+    evals_result = {}
+    params = {"objective": "binary", "metric": metric, "verbose": -1}
+    lgb.train(params, dtrain, num_boost_round=1, valid_sets=[dtrain], callbacks=[lgb.record_evaluation(evals_result)])
+    assert evals_result == {}
+
+
+def test_train_with_metric_python_none_uses_default_metric():
+    X, y = load_breast_cancer(return_X_y=True)
+    dtrain = lgb.Dataset(X, label=y)
+    dvalid = lgb.Dataset(X, label=y, reference=dtrain)
+    evals_result = {}
+    params = {"objective": "binary", "metric": None, "verbose": -1}
+    lgb.train(params, dtrain, num_boost_round=1, valid_sets=[dvalid], callbacks=[lgb.record_evaluation(evals_result)])
+    assert "binary_logloss" in evals_result["valid_0"]
+
+
 def test_quantized_training():
     X, y = make_synthetic_regression()
     ds = lgb.Dataset(X, label=y)

tests/python_package_test/test_sklearn.py

@@ -52,6 +52,7 @@
 
 SKLEARN_MAJOR, SKLEARN_MINOR, *_ = _sklearn_version.split(".")
 SKLEARN_VERSION_GTE_1_6 = (int(SKLEARN_MAJOR), int(SKLEARN_MINOR)) >= (1, 6)
+SKLEARN_VERSION_GTE_1_7 = (int(SKLEARN_MAJOR), int(SKLEARN_MINOR)) >= (1, 7)
 
 decreasing_generator = itertools.count(0, -1)
 estimator_classes = (lgb.LGBMModel, lgb.LGBMClassifier, lgb.LGBMRegressor, lgb.LGBMRanker)
@@ -453,7 +454,15 @@ def test_classifier_chain():
     X, y = make_multilabel_classification(n_samples=100, n_features=20, n_classes=n_outputs, random_state=0)
     X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.1, random_state=42)
     order = [2, 0, 1]
-    clf = ClassifierChain(base_estimator=lgb.LGBMClassifier(n_estimators=10), order=order, random_state=42)
+    # 'base_estimator' parameter was deprecated in scikit-learn 1.7 and removed in 1.9
+    #
+    #  * https://github.com/scikit-learn/scikit-learn/pull/30152
+    #  * https://github.com/scikit-learn/scikit-learn/pull/33750
+    #
+    if SKLEARN_VERSION_GTE_1_7:
+        clf = ClassifierChain(estimator=lgb.LGBMClassifier(n_estimators=10), order=order, random_state=42)
+    else:
+        clf = ClassifierChain(base_estimator=lgb.LGBMClassifier(n_estimators=10), order=order, random_state=42)
     clf.fit(X_train, y_train)
     score = clf.score(X_test, y_test)
     assert score >= 0.2
@@ -470,7 +479,15 @@ def test_regressor_chain():
     X, y = bunch["data"], bunch["target"]
     X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.1, random_state=42)
     order = [2, 0, 1]
-    reg = RegressorChain(base_estimator=lgb.LGBMRegressor(n_estimators=10), order=order, random_state=42)
+    # 'base_estimator' parameter was deprecated in scikit-learn 1.7 and removed in 1.9
+    #
+    #  * https://github.com/scikit-learn/scikit-learn/pull/30152
+    #  * https://github.com/scikit-learn/scikit-learn/pull/33750
+    #
+    if SKLEARN_VERSION_GTE_1_7:
+        reg = RegressorChain(estimator=lgb.LGBMRegressor(n_estimators=10), order=order, random_state=42)
+    else:
+        reg = RegressorChain(base_estimator=lgb.LGBMRegressor(n_estimators=10), order=order, random_state=42)
     reg.fit(X_train, y_train)
     y_pred = reg.predict(X_test)
     _, score, _ = mse(y_test, y_pred)