Datasets:
<unk> in Catalan corpus and Spanish sentences
Hello
Two small problems that I have observed.
1 - In the train_clean.ca there are occurences of the string, see:
"Condicions d'ús < unk > La teva privacitat"
This only ocurrs in the Catalan monolingual corpus. There are 9426 instances of it:
grep -rc "" train*
train_clean.ca:9426
train_clean.eu:0
Since is a well know token, it seems a tokenization problem.
2 - There are some Spanish sentences on the Catalan corpus, but I have not spent time in quantifying the issue. See an example:
"la Diputación Foral de Gipuzkoa, manifiestan en relación con la próxima convocatoria de promoción del personal auxiliar administrativo a plazas de Administrativo/a"
Thanks! Jordi
Hi Jordi! Thank you for the feedback. We will investigate and see where these errors have come from, those definitely shouldn't be present.
Hi Jordi,
Just to keep you in the loop - we have fixed the issue with the but are still investigating the presence of Spanish in the Catalan side of the corpus. We hope to have that solved soon.
Thanks so much. This is not urgent. I just wanted to make sure that you had visibility about this.
For visibility we use the corpus to train these two models:
https://huggingface.co/softcatala/translate-eus-cat
https://huggingface.co/softcatala/translate-cat-eus
Hello. Two additional quick comments:
- The new corpus published 5 days ago went from 9692996 to 10471139 strings. I guess that this is expected
- Consider deleting https://huggingface.co/datasets/projecte-aina/CA-EU_Parallel_Corpus/blob/main/train_clean.eu. This adds an additonal 1GB to the download
Best regards
Jordi,