diff --git a/diffsky/cosmos_utils/cosmos_mstar_model.py b/diffsky/cosmos_utils/cosmos_mstar_model.py
index fd2490f..f24e055 100644
--- a/diffsky/cosmos_utils/cosmos_mstar_model.py
+++ b/diffsky/cosmos_utils/cosmos_mstar_model.py
@@ -36,6 +36,29 @@
 def load_cosmos20_tdata(
     phot_keys=COSMOS_PHOT_KEYS, zlo=0.5, zhi=2.5, sm_key="lp_mass_best", **kwargs
 ):
+    """Load the COSMOS-20 dataset used to define training data for the scaling relation.
+
+    Parameters
+    ----------
+    phot_keys : list of strings
+
+    zlo, zhi : floats
+
+    sm_key : string
+
+    Returns
+    -------
+    photdata : namedtuple
+        Fields defined by PhotData at module top
+        Fields contain only the minimum information to train the approximate model
+
+    Notes
+    -----
+    This function has the following dependency:
+
+    https://github.com/LSSTDESC/cosmos20_colors
+
+    """
     try:
         from cosmos20_colors import load_cosmos20
     except (ImportError, ModuleNotFoundError):
@@ -67,6 +90,21 @@ def load_cosmos20_tdata(
 
 @jjit
 def predict_logsm(params, photdata):
+    """Approximate model of LePhare stellar mass from COSMOS griz photometry
+
+    Parameters
+    ----------
+    params : namedtuple
+        Fields defined by DEFAULT_PARAMS at top of module
+
+    photdata: namedtuple
+        Fields defined by PhotData at top of module
+
+    Returns
+    -------
+    logsm : ndarray, shape (n, )
+
+    """
     logsm = (
         params.b0
         + params.i * photdata.i
@@ -91,16 +129,33 @@ def _mae(x, y):
 
 
 @jjit
-def loss_kern(params, photdata):
+def _loss_kern(params, photdata):
     pred = predict_logsm(params, photdata)
     target = photdata.logsm
     return _mae(pred, target)
 
 
-loss_and_grad_func = jjit(value_and_grad(loss_kern))
+loss_and_grad_func = jjit(value_and_grad(_loss_kern))
 
 
 def fit_model(n_steps, photdata, params_init=DEFAULT_PARAMS, step_size=0.001):
+    """Find best-fitting parameters for the approximate stellar mass model
+
+    Parameters
+    ----------
+    n_steps : int
+
+    photdata: namedtuple
+        Fields defined by PhotData at top of module
+
+    Returns
+    -------
+    best_fit_params : namedtuple
+        Fields defined by DEFAULT_PARAMS at top of module
+
+    loss_arr : ndarray, shape (n_steps, )
+
+    """
     loss_collector = []
     opt_init, opt_update, get_params = jax_opt.adam(step_size)
     opt_state = opt_init(params_init)
diff --git a/diffsky/cosmos_utils/tests/test_cosmos_mstar_model.py b/diffsky/cosmos_utils/tests/test_cosmos_mstar_model.py
index dfa3591..eb32ac2 100644
--- a/diffsky/cosmos_utils/tests/test_cosmos_mstar_model.py
+++ b/diffsky/cosmos_utils/tests/test_cosmos_mstar_model.py
@@ -26,11 +26,11 @@ def test_fit_model():
         photdata = get_fake_cosmos_data(ran_key)
 
     logsm = cmm.predict_logsm(cmm.DEFAULT_PARAMS, photdata)
-    loss_init = cmm.loss_kern(cmm.DEFAULT_PARAMS, photdata)
+    loss_init = cmm._loss_kern(cmm.DEFAULT_PARAMS, photdata)
     assert np.all(np.isfinite(logsm))
     assert np.all(np.isfinite(loss_init))
 
     p_best, loss_arr = cmm.fit_model(100, photdata, step_size=0.01)
 
-    loss_best = cmm.loss_kern(p_best, photdata)
+    loss_best = cmm._loss_kern(p_best, photdata)
     assert loss_best < loss_init