diff --git a/src/batch_shuffle.rs b/src/batch_shuffle.rs
index df4b703..cfc72ee 100644
--- a/src/batch_shuffle.rs
+++ b/src/batch_shuffle.rs
@@ -13,7 +13,7 @@ use std::{
 
 use crate::{dataset::FSRSDataset, FSRSItem};
 
-pub struct BatchShuffledDataset<I> {
+pub(crate) struct BatchShuffledDataset<I> {
     dataset: Arc<FSRSDataset>,
     indices: Vec<usize>,
     input: PhantomData<I>,
diff --git a/src/cosine_annealing.rs b/src/cosine_annealing.rs
index 95751c0..325acf1 100644
--- a/src/cosine_annealing.rs
+++ b/src/cosine_annealing.rs
@@ -1,6 +1,6 @@
 use burn::{lr_scheduler::LrScheduler, tensor::backend::Backend, LearningRate};
 #[derive(Clone, Debug)]
-pub struct CosineAnnealingLR {
+pub(crate) struct CosineAnnealingLR {
     t_max: f64,
     eta_min: f64,
     init_lr: LearningRate,
diff --git a/src/dataset.rs b/src/dataset.rs
index 47cb3b2..82ec6e8 100644
--- a/src/dataset.rs
+++ b/src/dataset.rs
@@ -50,7 +50,7 @@ impl FSRSItem {
     }
 }
 
-pub struct FSRSBatcher<B: Backend> {
+pub(crate) struct FSRSBatcher<B: Backend> {
     device: B::Device,
 }
 
@@ -61,7 +61,7 @@ impl<B: Backend> FSRSBatcher<B> {
 }
 
 #[derive(Debug, Clone)]
-pub struct FSRSBatch<B: Backend> {
+pub(crate) struct FSRSBatch<B: Backend> {
     pub t_historys: Tensor<B, 2, Float>,
     pub r_historys: Tensor<B, 2, Float>,
     pub delta_ts: Tensor<B, 1, Float>,
@@ -133,7 +133,7 @@ impl<B: Backend> Batcher<FSRSItem, FSRSBatch<B>> for FSRSBatcher<B> {
     }
 }
 
-pub struct FSRSDataset {
+pub(crate) struct FSRSDataset {
     items: Vec<FSRSItem>,
 }
 
diff --git a/src/inference.rs b/src/inference.rs
index ad2d57e..4d5b64f 100644
--- a/src/inference.rs
+++ b/src/inference.rs
@@ -13,10 +13,10 @@ use crate::model::Model;
 use crate::training::BCELoss;
 use crate::{FSRSError, FSRSItem};
 use burn::tensor::ElementConversion;
-pub const DECAY: f64 = -0.5;
+pub(crate) const DECAY: f64 = -0.5;
 /// (9/10) ^ (1 / DECAY) - 1
-pub const FACTOR: f64 = 19f64 / 81f64;
-pub const S_MIN: f32 = 0.01;
+pub(crate) const FACTOR: f64 = 19f64 / 81f64;
+pub(crate) const S_MIN: f32 = 0.01;
 /// This is a slice for efficiency, but should always be 17 in length.
 pub type Parameters = [f32];
 use itertools::izip;
diff --git a/src/model.rs b/src/model.rs
index 19b7f16..08ef3cd 100644
--- a/src/model.rs
+++ b/src/model.rs
@@ -16,7 +16,7 @@ pub struct Model<B: Backend> {
     pub config: ModelConfig,
 }
 
-pub trait Get<B: Backend, const N: usize> {
+pub(crate) trait Get<B: Backend, const N: usize> {
     fn get(&self, n: usize) -> Tensor<B, N>;
 }
 
@@ -177,7 +177,7 @@ impl<B: Backend> Model<B> {
 }
 
 #[derive(Debug, Clone)]
-pub struct MemoryStateTensors<B: Backend> {
+pub(crate) struct MemoryStateTensors<B: Backend> {
     pub stability: Tensor<B, 1>,
     pub difficulty: Tensor<B, 1>,
 }
@@ -240,7 +240,7 @@ impl<B: Backend> FSRS<B> {
     }
 }
 
-pub fn parameters_to_model<B: Backend>(parameters: &Parameters) -> Model<B> {
+pub(crate) fn parameters_to_model<B: Backend>(parameters: &Parameters) -> Model<B> {
     let config = ModelConfig::default();
     let mut model = Model::new(config);
     model.w = Param::from(Tensor::from_floats(
diff --git a/src/pre_training.rs b/src/pre_training.rs
index a3ae59a..872c1dc 100644
--- a/src/pre_training.rs
+++ b/src/pre_training.rs
@@ -106,7 +106,7 @@ fn loss(
     logloss + l1
 }
 
-pub const INIT_S_MAX: f32 = 100.0;
+pub(crate) const INIT_S_MAX: f32 = 100.0;
 
 fn search_parameters(
     mut pretrainset: HashMap<FirstRating, Vec<AverageRecall>>,
diff --git a/src/training.rs b/src/training.rs
index 65a57d5..90bf52d 100644
--- a/src/training.rs
+++ b/src/training.rs
@@ -166,7 +166,7 @@ impl MetricsRenderer for ProgressCollector {
 }
 
 #[derive(Config)]
-pub struct TrainingConfig {
+pub(crate) struct TrainingConfig {
     pub model: ModelConfig,
     pub optimizer: AdamConfig,
     #[config(default = 5)]
diff --git a/src/weight_clipper.rs b/src/weight_clipper.rs
index 0c475ec..13e2df0 100644
--- a/src/weight_clipper.rs
+++ b/src/weight_clipper.rs
@@ -4,7 +4,7 @@ use crate::{
 };
 use burn::tensor::{backend::Backend, Data, Tensor};
 
-pub fn weight_clipper<B: Backend>(parameters: Tensor<B, 1>) -> Tensor<B, 1> {
+pub(crate) fn weight_clipper<B: Backend>(parameters: Tensor<B, 1>) -> Tensor<B, 1> {
     let val = clip_parameters(&parameters.to_data().convert().value);
     Tensor::from_data(
         Data::new(val, parameters.shape()).convert(),
@@ -12,7 +12,7 @@ pub fn weight_clipper<B: Backend>(parameters: Tensor<B, 1>) -> Tensor<B, 1> {
     )
 }
 
-pub fn clip_parameters(parameters: &Parameters) -> Vec<f32> {
+pub(crate) fn clip_parameters(parameters: &Parameters) -> Vec<f32> {
     // https://regex101.com/r/21mXNI/1
     const CLAMPS: [(f32, f32); 17] = [
         (S_MIN, INIT_S_MAX),