diff --git a/benches/cakes/src/steps.rs b/benches/cakes/src/steps.rs
index 5ad362dc..96e93f55 100644
--- a/benches/cakes/src/steps.rs
+++ b/benches/cakes/src/steps.rs
@@ -9,7 +9,7 @@ use abd_clam::{
     metric::ParMetric,
     Ball, Cluster, Dataset, FlatVec,
 };
-use bench_utils::{ann_benchmarks::AnnDataset, reports::CakesResults, types::Row};
+use bench_utils::{ann_benchmarks::AnnDataset, reports::CakesResults};
 use distances::Number;
 use rand::prelude::*;
 
@@ -421,7 +421,7 @@ pub fn read_tabular<P: AsRef<std::path::Path>>(
     seed: Option<u64>,
     inp_dir: &P,
     out_dir: &P,
-) -> Result<(Vec<Row<f32>>, Vec<Vec<(usize, f32)>>), String> {
+) -> Result<(Vec<Vec<f32>>, Vec<Vec<(usize, f32)>>), String> {
     let data_path = out_dir.as_ref().join(format!("{}-0.flat_vec", dataset.name()));
     let queries_path = out_dir.as_ref().join(format!("{}.queries", dataset.name()));
     let neighbors_path = out_dir.as_ref().join(format!("{}.neighbors", dataset.name()));
@@ -443,29 +443,49 @@ pub fn read_tabular<P: AsRef<std::path::Path>>(
         let (train, queries, neighbors) = (data.train, data.queries, data.neighbors);
         let (min_dim, max_dim) = train
             .iter()
+            .chain(queries.iter())
             .fold((usize::MAX, 0), |(min, max), x| (min.min(x.len()), max.max(x.len())));
 
-        let items = train.iter().cloned().map(Row::from).collect();
-        let data = FlatVec::new(items)?
+        let data = FlatVec::new(train)?
             .with_name(dataset.name())
             .with_dim_lower_bound(min_dim)
             .with_dim_upper_bound(max_dim);
         ftlog::info!("Writing original data to {data_path:?}...");
         data.write_to(&data_path)?;
 
-        let queries = queries.iter().cloned().map(Row::from).collect::<Vec<_>>();
+        let npy_name = format!("{}-0.npy", dataset.name());
+        ftlog::info!(
+            "Writing original data as npy array to {:?}/{npy_name}...",
+            out_dir.as_ref()
+        );
+        data.write_npy(out_dir, &npy_name)?;
+
+        let query_data = FlatVec::new(queries)?
+            .with_name(&format!("{}-queries", dataset.name()))
+            .with_dim_lower_bound(min_dim)
+            .with_dim_upper_bound(max_dim);
+        let query_npy_name = format!("{}.npy", query_data.name());
+        ftlog::info!(
+            "Writing queries as npy array to {:?}/{query_npy_name}...",
+            out_dir.as_ref()
+        );
+        query_data.write_npy(out_dir, &query_npy_name)?;
+        let queries = query_data.take_items();
+
         std::fs::write(&queries_path, bitcode::encode(&queries).map_err(|e| e.to_string())?)
             .map_err(|e| e.to_string())?;
         std::fs::write(&neighbors_path, bitcode::encode(&neighbors).map_err(|e| e.to_string())?)
             .map_err(|e| e.to_string())?;
 
-        let base_cardinality = data.cardinality();
+        let train = data.take_items();
+        let base_cardinality = train.len();
         let data = AnnDataset {
             train,
             queries: Vec::new(),
             neighbors: Vec::new(),
         }
         .augment(1 << max_power, 0.1);
+
         let data = FlatVec::new(data.train)?
             .with_dim_lower_bound(min_dim)
             .with_dim_upper_bound(max_dim);
@@ -477,6 +497,14 @@ pub fn read_tabular<P: AsRef<std::path::Path>>(
             let data_path = out_dir.as_ref().join(format!("{name}.flat_vec"));
             ftlog::info!("Writing {}x augmented data to {data_path:?}...", 1 << power);
             data_sample.write_to(&data_path)?;
+
+            let npy_name = format!("{name}.npy");
+            ftlog::info!(
+                "Writing {}x augmented data as npy array to {:?}/{npy_name}...",
+                1 << power,
+                out_dir.as_ref()
+            );
+            data_sample.write_npy(out_dir, &npy_name)?;
         }
 
         Ok((queries, neighbors))
@@ -489,7 +517,7 @@ pub fn read_member_set<P: AsRef<std::path::Path>>(
     dataset: &bench_utils::ann_benchmarks::RawData,
     inp_dir: &P,
     out_dir: &P,
-) -> Result<(std::path::PathBuf, Vec<Row<usize>>, Vec<Vec<(usize, f32)>>), String> {
+) -> Result<(std::path::PathBuf, Vec<Vec<usize>>, Vec<Vec<(usize, f32)>>), String> {
     let data_name = format!("{}.flat_vec", dataset.name());
     let data_path = out_dir.as_ref().join(&data_name);
     let queries_path = out_dir.as_ref().join(format!("{}.queries", dataset.name()));
@@ -511,12 +539,10 @@ pub fn read_member_set<P: AsRef<std::path::Path>>(
             .iter()
             .fold((usize::MAX, 0), |(min, max), x| (min.min(x.len()), max.max(x.len())));
 
-        let data = train.into_iter().map(Row::from).collect();
-        let data = FlatVec::new(data)?
+        let data = FlatVec::new(train)?
             .with_name(&data_name)
             .with_dim_lower_bound(min_dim)
             .with_dim_upper_bound(max_dim);
-        let queries = queries.into_iter().map(Row::from).collect::<Vec<_>>();
 
         ftlog::info!("Writing data to {data_path:?}...");
         data.write_to(&data_path)?;
diff --git a/crates/abd-clam/src/core/dataset/flat_vec.rs b/crates/abd-clam/src/core/dataset/flat_vec.rs
index 742d9a4e..5585cfe8 100644
--- a/crates/abd-clam/src/core/dataset/flat_vec.rs
+++ b/crates/abd-clam/src/core/dataset/flat_vec.rs
@@ -193,6 +193,12 @@ impl<I, Me> FlatVec<I, Me> {
         &self.items
     }
 
+    /// Takes the items out of the dataset.
+    #[must_use]
+    pub fn take_items(self) -> Vec<I> {
+        self.items
+    }
+
     /// Transforms the items in the dataset.
     ///
     /// # Type Parameters