scikit-learn · ogrisel · Jun 29, 2022 · Jun 13, 2022 · Jun 13, 2022 · Jun 14, 2022
diff --git a/doc/whats_new/v1.2.rst b/doc/whats_new/v1.2.rst
@@ -30,6 +30,12 @@ random sampling procedures.
   :pr:`10805` by :user:`Mathias Andersen <MrMathias>` and
   :pr:`23471` by :user:`Meekail Zain <micky774>`
 
+- |Enhancement| :class:`linear_model.GammaRegressor`,
+  :class:`linear_model.PoissonRegressor` and :class:`linear_model.TweedieRegressor`
+  can reach higher precision with the lbfgs solver, in particular when `tol` is set
+  to a tiny value. Moreover, `verbose` is now properly propagated to L-BFGS-B.
+  :pr:`23619` by :user:`Christian Lorentzen <lorentzenchr>`.
+
 Changes impacting all modules
 -----------------------------
 
@@ -146,6 +152,12 @@ Changelog
   `solver="newton-cg"`, `fit_intercept=True`, and a single feature. :pr:`23608`
   by `Tom Dupre la Tour`_.
 
+- |Enhancement| :class:`linear_model.GammaRegressor`,
+  :class:`linear_model.PoissonRegressor` and :class:`linear_model.TweedieRegressor`
+  can reach higher precision with the lbfgs solver, in particular when `tol` is set
+  to a tiny value. Moreover, `verbose` is now properly propagated to L-BFGS-B.
+  :pr:`23619` by :user:`Christian Lorentzen <lorentzenchr>`.
+
 - |API| The default value for the `solver` parameter in
   :class:`linear_model.QuantileRegressor` will change from `"interior-point"`
   to `"highs"` in version 1.4.
@@ -158,7 +170,7 @@ Changelog
   negative likelihood ratios derived from the confusion matrix
   of a binary classification problem. :pr:`22518` by
   :user:`Arturo Amor <ArturoAmorQ>`.
-  
+
 - |Fix| :func:`metrics.ndcg_score` will now trigger a warning when the `y_true`
   value contains a negative value. Users may still use negative values, but the
   result may not be between 0 and 1. Starting in v1.4, passing in negative

diff --git a/sklearn/linear_model/_glm/glm.py b/sklearn/linear_model/_glm/glm.py
@@ -285,9 +285,13 @@ def fit(self, X, y, sample_weight=None):
                 jac=True,
                 options={
                     "maxiter": self.max_iter,
-                    "iprint": (self.verbose > 0) - 1,
+                    "maxls": 50,  # default is 20
+                    "iprint": self.verbose - 1,
                     "gtol": self.tol,
-                    "ftol": 1e3 * np.finfo(float).eps,
+                    # The constant 64 was found empirically to pass the test suite.
+                    # The point is that ftol is very small, but a bit larger than
+                    # machine precision for float64, which is the dtype used by lbfgs.
+                    "ftol": 64 * np.finfo(float).eps,
                 },
                 args=(X, y, sample_weight, l2_reg_strength, n_threads),
             )