Datasets:

CrocoSign
/

pjm-data

texts_fixed.csv holds the Polish and English text of the 29,928 dataset pairs (the keys in splits/split_*_ms.csv and splits/split_*_filtered.csv), each in two versions:

pl, en: the texts as first published. All results in the CrocoSign paper use en as the reference.
pl_fixed, en_fixed: corrected versions. fixed is 1 when either text differs from the original.

The original files in translations/ and splits/ are unchanged.

What was corrected (2026-10-31):

Polish, 211 pairs: typos and missing diacritics (ze for że, Pam for Pan), and cut-off word fragments, removed while keeping the signer's self-correction (słoik z kot... żabą becomes słoik z żabą).
English, 1,440 pairs (4.8%): meaning errors of the machine translation. Most are systematic: migać rendered as "blink" or "flash" instead of "sign", wskazówka (clock hand) as "pointer" or "clue", kosz (basket) as "garbage can", kula (ball) as "bullet", smarkacz (brat) as "snot" or "greaser", Pan as "The Lord", drezyna (draisine) as "dreidel", and akurat left untranslated. Others follow from a Polish typo, such as "the stove" for Piec, a typo of Pies (dog).

In all, 1,547 pairs have fixed = 1 (211 with a Polish change, 1,440 with an English change; some have both). The English corrections were made against pl_fixed.

How we found them: a dictionary scan (hunspell, Morfeusz 2) and a language model that read every Polish sentence and then every Polish-English pair and proposed corrections. We approved the Polish list and spot-checked the English one (25 random flags, all real errors), and checked recall on the systematic terms above; the cases the model missed there (40, mostly migać) and 5 clock times were corrected by hand. Other errors that the model did not flag remain. The Polish keeps the signers' colloquial wording, and forms such as na dworzu are left as they are.

import pandas as pd
texts = pd.read_csv("texts/texts_fixed.csv").set_index("key")
target = texts["en_fixed"]  # or texts["en"] to reproduce the paper