Remove ---- .... === ### dupes from datasets and tokenizer

#3
by foss22 - opened

Remove multiple ---- .... === ### dupes from datasets and tokenizer.

Capacity ot the limited weights and compute should not be spent on learning the
difference of .... and ...
difference of ---- and --------
difference of ======== and ================
difference of #### and #####

Better spend this capacity on the difference of words and subwords. Prioritize meaning over formatting garbage.

ama-77/wd_0$ rg -v [A-Za-z0-9] tokenizer.json|sed 's/,$//g' |huniq -c|rg \"|ug -v [A-Za-z]|ug  '=|--|\#|\.\.'|sort -rnk1|cat -n

     1  30         "--"
     2  14         "="
     3  11         "--------"
     4  11         "----"
     5  7         "#"
     6  6         ".."
     7  5         "...."
     8  4         "..."
     9  4         "----------------"
    10  4         "------"
    11  3         "===="
    12  3         "=="
    13  3         ".="
    14  3         "........"
    15  3         "-----+"
    16  3         "----+"
    17  3         "---+"
    18  3         "+--------"
    19  2         "================"
    20  2         "========"
    21  2         ".--"
    22  2         "--\""
    23  2         "--------------------------------"
    24  2         "------+"
    25  2         "-----+-----+"
    26  2         "---+---+"
    27  2         "--+"
    28  2         "##"
    29  1         "Ġ="
    30  1         "Ġ.."
    31  1         "Ġ----"
    32  1         "Ġ--"
    33  1         "_#"
    34  1         "................"
    35  1         "--_"
    36  1         "----------------------------------------------------------------"
    37  1         "-------+"
    38  1         ",--"
    39  1         "#|"
    40  1         "#####|"
    41  1         "####"

How many for removal?

time for f in 000{9..0}.parquet;do arrowcat $f;done|rg -c -- --
62346

real    0m13,305s

That's a good point. Weirdly enough I built this tokenizer mostly from Gutenberg so it's a bit weird.
It is too late now to change it for the current models, because I'm already training and experimenting with some runs, but I'll consider building a token-v4.
Thank you

Sign up or log in to comment