diff --git a/rslib/src/scheduler/fsrs/retention.rs b/rslib/src/scheduler/fsrs/retention.rs
index 07602a09c..f4dc1935d 100644
--- a/rslib/src/scheduler/fsrs/retention.rs
+++ b/rslib/src/scheduler/fsrs/retention.rs
@@ -135,15 +135,14 @@ impl Collection {
                     r.review_kind == RevlogReviewKind::Review
                         && r.button_chosen > 0
                         && r.taken_millis > 0
+                        && r.taken_millis < 1200000 // 20 minutes
                 })
                 .sorted_by(|a, b| a.button_chosen.cmp(&b.button_chosen))
                 .group_by(|r| r.button_chosen)
                 .into_iter()
                 .for_each(|(button_chosen, group)| {
                     let group_vec = group.into_iter().map(|r| r.taken_millis).collect_vec();
-                    let average_secs =
-                        group_vec.iter().sum::<u32>() as f64 / group_vec.len() as f64 / 1000.0;
-                    arr[button_chosen as usize - 1] = average_secs
+                    arr[button_chosen as usize - 1] = median_secs(&group_vec);
                 });
             if arr == default {
                 return Err(AnkiError::FsrsInsufficientData);
@@ -157,19 +156,23 @@ impl Collection {
                     r.review_kind == RevlogReviewKind::Learning
                         && r.button_chosen >= 1
                         && r.taken_millis > 0
+                        && r.taken_millis < 1200000 // 20 minutes
                 })
                 .map(|r| r.taken_millis);
-            let length = revlogs_filter.clone().count() as f64;
-            if length > 0.0 {
-                revlogs_filter.sum::<u32>() as f64 / length / 1000.0
-            } else {
-                return Err(AnkiError::FsrsInsufficientData);
-            }
+            let group_vec = revlogs_filter.collect_vec();
+            median_secs(&group_vec)
         };
+        if learn_cost == 0.0 {
+            return Err(AnkiError::FsrsInsufficientData);
+        }
 
         let forget_cost = {
             let review_kind_to_total_millis = revlogs
                 .iter()
+                .filter(|r| {
+                    r.button_chosen > 0 && r.taken_millis > 0 && r.taken_millis < 1200000
+                    // 20 minutes
+                })
                 .sorted_by(|a, b| a.cid.cmp(&b.cid).then(a.id.cmp(&b.id)))
                 .group_by(|r| r.review_kind)
                 /*
@@ -192,12 +195,7 @@ impl Collection {
             }
             let mut arr = [0.0; 5];
             for (review_kind, group) in group_sec_by_review_kind.iter().enumerate() {
-                let average_secs = group.iter().sum::<u32>() as f64 / group.len() as f64 / 1000.0;
-                arr[review_kind] = if average_secs.is_nan() {
-                    0.0
-                } else {
-                    average_secs
-                }
+                arr[review_kind] = median_secs(group);
             }
             arr
         };
@@ -221,3 +219,20 @@ impl Collection {
         Ok(params)
     }
 }
+
+fn median_secs(group: &Vec<u32>) -> f64 {
+    let length = group.len();
+    if length > 0 {
+        let mut group_vec = group.clone();
+        group_vec.sort_unstable();
+        let median_millis = if length % 2 == 0 {
+            let mid = length / 2;
+            (group_vec[mid - 1] + group_vec[mid]) as f64 / 2.0
+        } else {
+            group_vec[length / 2] as f64
+        };
+        median_millis / 1000.0
+    } else {
+        0.0
+    }
+}