diff --git a/docling_eval/dataset_builders/cvat_dataset_builder.py b/docling_eval/dataset_builders/cvat_dataset_builder.py
index 0ca8d51c..d7d4d08c 100644
--- a/docling_eval/dataset_builders/cvat_dataset_builder.py
+++ b/docling_eval/dataset_builders/cvat_dataset_builder.py
@@ -379,7 +379,188 @@ def parse_annotation(self, image_annot: Dict) -> Tuple:
             merges,
             group,
         )
+    
+    def _validate_annotations_consistency(self,
+                                          boxes,
+                                          lines,
+                                          reading_order,
+                                          to_captions,
+                                          to_footnotes,
+                                          to_values,
+                                          merges,
+                                          groups) -> bool:
+        """Validate annotations consistency."""
+
+        def _validate_to_captions():
+            """validate to captions.
+
+            - Make sure the first bbox is a FloatingItem and the subsequent
+            are of type caption.
+            - Make sure that the to_caption only starts at the beginning of
+            a group, if a group is present.
+            """
+            for to_caption in to_captions:
+
+                boxids = to_caption["boxids"]
+                if len(boxids)!=2:
+                    _log.error(f"to_caption is longer than 2: {len(to_caption['boxids'])}")
+                    return False
+
+                box_source = boxes[boxids[0]]
+                box_target = boxes[boxids[1]]
+
+                label_source = DocItemLabel(box_source["@label"])
+                label_target = DocItemLabel(box_target["@label"])
+
+                if label_source not in [
+                        DocItemLabel.PICTURE,
+                        DocItemLabel.TABLE,
+                        DocItemLabel.CODE,
+                        DocItemLabel.FORM,
+                ]:
+                    _log.error(f"to_caption label-source: {label_source}")
+                    return False
+                
+                if label_target!=DocItemLabel.CAPTION:
+                    _log.error(f"to_caption label-target: {label_target}")
+                    return False
+                
+            return True
+        
+        def _validate_to_footnotes():
+            """validate to footnotes.
+
+            - Make sure the first bbox is a FloatingItem and the subsequent
+            are of type footnote.
+            - Make sure that the to_footnote only starts at the beginning of
+            a group, if a group is present.
+            """
+            for to_footnote in to_footnotes:
+
+                boxids = to_footnote["boxids"]
+                if len(boxids)!=2:
+                    _log.error(f"to_footnote is longer than 2: {len(to_footnote['boxids'])}")
+                    return False
+
+                box_source = boxes[boxids[0]]
+                box_target = boxes[boxids[1]]
+
+                label_source = DocItemLabel(box_source["@label"])
+                label_target = DocItemLabel(box_target["@label"])
+
+                if label_source not in [
+                        DocItemLabel.PICTURE,
+                        DocItemLabel.TABLE,
+                        DocItemLabel.CODE,
+                        DocItemLabel.FORM,
+                ]:
+                    _log.error(f"to_footnote label-source: {label_source}")
+                    return False
+                
+                if label_target!=DocItemLabel.FOOTNOTE:
+                    _log.error(f"to_footnote label-target: {label_target}")
+                    return False
+            
+            return True        
+
+        def _validate_captions():
+            """validate captions.
+
+            Make sure each caption either is the start of a to_caption or
+            has a merge node.
+            """
+            caption_ids = []            
+            for to_caption in to_captions:
+                boxids = to_caption["boxids"]
+                if len(boxids)!=2:
+                    _log.error(f"to_caption is longer than 2: {len(to_caption['boxids'])}")
+                    return False
+
+                box_target = boxes[boxids[1]]
+                label_target = DocItemLabel(box_target["@label"])
+
+                if label_target==DocItemLabel.CAPTION:
+                    caption_ids.append(boxids[1])
+            
+            for i,box in enumerate(boxes):
+                label = DocItemLabel(box_source["@label"])
+
+                caption_is_linked = False
+                if label == DocItemLabel.CAPTION:
+                    if i not in caption_ids:
+                        caption_is_linked = True
+
+                if not caption_is_linked:
+                    _log.error(f"caption {box} is not linked to anything")
+                    return False
+                    
+            return True
+        
+        def _validate_list_items():
+            """validate list items.
+
+            Make sure every list-item has a group or merge point.
+            """
+            for i,box in enumerate(boxes):
+                label = DocItemLabel(box_source["@label"])
+
+                listitem_is_grouped = True
+                if label == DocItemLabel.LIST_ITEM:
+                    for group in groups:
+                        if i in group["boxids"]:
+                            listitem_is_grouped = True
+
+                listitem_is_merged = True
+                if label == DocItemLabel.LIST_ITEM:
+                    for merge in merges:
+                        if i in merge["boxids"]:
+                            listitem_is_merged = True
+                            
+                if (not listitem_is_linked) or (not listitem_is_merged):
+                    _log.error(f"listitem {box} is not grouped or merged.")
+                    return False
+
+            return True
+        
+        def _validate_reading_order():
+            """validate reading-order.
+
+            Make sure that every bbox has a reading-order point or is
+            entirely overlapped with a bbox that has a reading-order. 
+            """
+            return True
+            
+        def _validate_merges():
+            """validate merges.
+
+            Make sure that all bbox in a merge are of same type
+            """
+            return True        
+
+        def _validate_group():
+            """validate groups.
 
+            Make sure that all bbox in a group are of same type
+            """
+            return True        
+
+        def _validate_to_values_links():
+            """validate to_value links.
+
+            Make sure that all to_value links are between
+            """
+            return True        
+        
+
+        return (
+            _validate_to_captions() and
+            _validate_to_footnotes() and
+            _validate_reading_order() and
+            _validate_merges() and
+            _validate_group() and
+            __validate_list_items()
+        )
+    
     def create_prov(
         self,
         box: Dict,
@@ -895,6 +1076,17 @@ def create_true_document(
             _log.error(f"Incorrect annotation for {basename}: no reading order found")
             return None
 
+        if self._validate_annotations_consistency(boxes=boxes,
+                                                  lines=lines,
+                                                  reading_order=reading_order,
+                                                  to_captions=to_captions,
+                                                  to_footnotes=to_footnotes,
+                                                  to_values=to_values,
+                                                  merges=merges,
+                                                  group=group):
+            _log.error(f"Inconsistent annotation for {basename}: skipping")
+            return None
+        
         # Original Groundtruth and Prediction files
         orig_file = desc.document_file