diff --git a/jack/readers/extractive_qa/util.py b/jack/readers/extractive_qa/util.py
index 46402d8c..d9c6e44c 100644
--- a/jack/readers/extractive_qa/util.py
+++ b/jack/readers/extractive_qa/util.py
@@ -109,19 +109,22 @@ def prepare_data(qa_setting: QASetting,
                     max_answer = max(max_answer, end)
 
         # cut support whenever there is a maximum allowed length and recompute answer spans
-        if max_support_length is not None and len(support_tokens) > max_support_length > 0:
+        support_length = all_support_length[doc_idx]
+        if max_support_length is not None and support_length > max_support_length > 0:
             if max_answer < max_support_length:
                 # Find new start and end in the flattened support
                 new_start = 0
                 new_end = max_support_length
             else:
                 offset = rng.randint(1, 11)
-                new_end = max_answer + offset
-                new_start = max(0, min(min_answer - offset, new_end - max_support_length))
-                while new_end - new_start > max_support_length:
-                    answer_spans = [(s, e) for s, e in answer_spans if e < (new_end - offset)]
-                    new_end = max(answer_spans, key=lambda span: span[1])[1] + offset
-                    new_start = max(0, min(min_answer - offset, new_end - max_support_length))
+                new_end = max_answer
+                new_start = max(0, min(min_answer, new_end + 2 * offset - max_support_length))
+                while new_end - new_start > max_support_length - 2 * offset:
+                    answer_spans = [(s, e) for s, e in answer_spans if e < new_end]
+                    new_end = max(answer_spans, key=lambda span: span[1])[1]
+                    new_start = max(0, min(min_answer, new_end + 2 * offset - max_support_length))
+                new_end = min(new_end + offset, support_length)
+                new_start = max(new_start - offset, 0)
 
             # Crop support according to new start and end pointers
             all_support_tokens[doc_idx] = support_tokens[new_start:new_end]