Gold check: 41 identical review texts carry both polarity labels (40 train, 1 test — all byte-identical, 106 rows)

#5
by nootxlm - opened

I grouped the amazon_polarity config by the review text (content) and counted texts that carry more than one label, over the full census of 4,000,000 rows. An independent byte-exact recount with a separate script reproduced both counts exactly:

  • train (3,600,000 rows): 40 byte-identical texts with two labels — 104 rows
  • test (400,000 rows): 1 byte-identical text with two labels — 2 rows

The test pair is the sharp case. The same 381-character review body ("This tool is absolutely fabulous for doing the few things you will need it for, but and this is a BIG but…") appears twice in test, once with label 1 under the title "Great plunge saw, but..." and once with label 0 under the title "not a suitable tool for this application" (the two copies hash identically). A model scored on the content field is marked wrong on one copy whichever label it gives.

The train pairs have the same shape at larger count: short review bodies repeated across products, carrying both labels (one text appears 8 times with both labels among its copies). Reported at exactly this size: 41 texts, 106 rows, in a 4,000,000-row census.

Sign up or log in to comment