fix RAM issue

Don't add the binary to the dictionaries in `annotations`. Instead, use a copy of the dictionary that can be garbage collected after each `yield`

DUDE_loader.py

@@ -222,9 +222,10 @@ class DUDE(datasets.GeneratorBasedBuilder):
         annotations = [x for x in annotations if x["data_split"] == split]
 
         for i, a in enumerate(annotations):
-            a["data_split"] = split
             if a["docId"] in SKIP_DOC_IDS:
                 continue
+            a = dict(a)
+            a["data_split"] = split
             a["answers_page_bounding_boxes"] = parse_bbox(a["answers_page_bounding_boxes"])
             docpath = retrieve_doc(a["docId"])
             ocrpath = retrieve_OCR(a["docId"])