Each AMR dataset (e.g. the dev set of LDC2017T10) consists of multiple files. In preprocessing, these files are concatenated in arbitrary order (determined by the OS). This order is not always consistent (I just had a case where one file from a first preprocessing run and a file from a different preprocessing run had different orders, leading to problems).
The concatenation should instead be consistent over different runs, e.g. by lexical ordering of filenames.
Each AMR dataset (e.g. the dev set of LDC2017T10) consists of multiple files. In preprocessing, these files are concatenated in arbitrary order (determined by the OS). This order is not always consistent (I just had a case where one file from a first preprocessing run and a file from a different preprocessing run had different orders, leading to problems).
The concatenation should instead be consistent over different runs, e.g. by lexical ordering of filenames.