diff --git a/sae_training/evals.py b/sae_training/evals.py
index 276608e1..c06017fa 100644
--- a/sae_training/evals.py
+++ b/sae_training/evals.py
@@ -26,10 +26,15 @@ def run_evals(
     eval_tokens = activation_store.get_batch_tokens()
 
     # Get Reconstruction Score
-    recons_score, ntp_loss, recons_loss, zero_abl_loss = get_recons_loss(
-        sparse_autoencoder, model, activation_store
+    losses_df = recons_loss_batched(
+        sparse_autoencoder, model, activation_store, n_batches = 10,
     )
 
+    recons_score = losses_df["score"].mean()
+    ntp_loss = losses_df["loss"].mean()
+    recons_loss = losses_df["recons_loss"].mean()
+    zero_abl_loss = losses_df["zero_abl_loss"].mean()
+
     # get cache
     _, cache = model.run_with_cache(
         eval_tokens,
@@ -144,8 +149,9 @@ def head_replacement_hook(activations, hook):
 def recons_loss_batched(sparse_autoencoder, model, activation_store, n_batches=100):
     losses = []
     for _ in tqdm(range(n_batches)):
+        batch_tokens = activation_store.get_batch_tokens()
         score, loss, recons_loss, zero_abl_loss = get_recons_loss(
-            sparse_autoencoder, model, activation_store
+            sparse_autoencoder, model, batch_tokens
         )
         losses.append(
             (