When using FARM via Haystack for open domain QA, I am often getting the following ERROR

ERROR - farm.modeling.predictions -   Both start and end offsets should be 0: 1053, 1053 with a no_answer. 

On closer inspection, it seems not all our QACandidates are being converted to document level. I suspect the issue lies in aggregate_preds( )

preds_d = self.aggregate_preds(preds, passage_start_t, ids, seq_2_start_t)

This means that in our list of final predictions, we are getting a mix of no_answer QACandidates: some with QACandidate.aggregation_level = 'passage' where start = end = 1053 for example and some with QACandidate.aggregation_level = 'document' where start = end = 0.