I see default params for sacrebleu as smooth_method=exp and tokenize=None (uses 13a default tokenizer) while in sacrebleu python library (GitHub - mjpost/sacrebleu: Reference BLEU implementation that auto-downloads test sets and reports a version string to facilitate cross-lab comparisons) , the default params seem to be smooth_method=floor and tokenize='none' (which uses BaseTokenizer).
Which set is officially recognized and why was exp smooth method chosen in hf datasets ?
1 Like