Yoruba seems odd.
Looking at the smolsent/en_yo.jsonl the text seems odd. I asked GPT (multiple models), and they responded with example these could be the issues:
- Diacritics mistakes – missing or incorrect accents.
- Uncommon/obscure words – confusing for readers.
- Literal translation effects – some phrases sound unnatural in Yoruba.
I'm not sure about any of that since I'm not a native speaker, but as someone who sees different languages routinely, it looks different from the classic Yoruba I've seen in other datasets (and even from smoldoc).
"I asked GPT" 🤢🤢
@menacing-american-patriot I did not want to disclose the exact names of the models I used, so I just said “GPT.” I asked multiple models. I believe they are more knowledgeable in that language than I am. It was easy for them to find the diacritic mistakes with examples, and I could validate them against sources such as Wikipedia.
Are you a Yoruba speaker or a smol maintainer?
Now say “sources such as Wikipedia” 🤢🤢