No, it's not bad data. In addition to the non-AUG start, it could be because they are annotated as CDS 5' incomplete and the start was therefore left open. See this example. CDS 5' (or CDS 3') incomplete transcripts are manually annotated by the HAVANA team and displayed in Ensembl as part of the GENCODE gene set.
You can also check the fasta deflines in the file and specifically the status and transcript types. I count 58 CDS or 0.25% using the ensembl_havana_transcript:known. And here's some R code with details.