Datasets:

Modalities:
Tabular
Text
Formats:
parquet
ArXiv:
Libraries:
Datasets
Dask
License:

Incosistency in Greek data

#2
by gpantaz - opened

Hello!

Many thanks for this, it is really cool! I am a native Greek speaker and I have noticed there is a repeating mistake in the input and output responses. Specifically, the insertion of the word Ασφάλεια (which means safety / security depending on context) but it generally feels that it is misplaced. For example row 5 contains the response Ασφάλεια Χαρυ.

I tried running the nllb-200-distilled-3.3B model (I assume you are using this based on the Aya paper.) on the english prompt / response and I am getting this:

>>> translator = pipeline("translation", model="facebook/nllb-200-3.3B", device="cpu", src_lang="eng_Latn", tgt_lang="ell_Grek")
>>> translator("What are the answers to this following set of questions: 1. who had a farm ? Numbered answers: 1. Harry")
>>>[{'translation_text': 'Ποιες είναι οι απαντήσεις στα ακόλουθα ερωτήματα: 1. ποιος είχε φάρμα;'}

Happy to help if I can to solve this.
George

Cohere For AI org
edited Mar 15

Hi @gpantaz , thanks for bringing this to our notice. The translations from the NLLB model were not always good.
Let me check on this and get back to you shortly!

Cohere For AI org

Hello @gpantaz , thanks for flagging this! I took a look into this and as it turns out this is an error due to our sentence splitting package. NLLB handles newlines very inconsistently so all our inputs are split into sentences first then pieced back together. As you can see Coqa has a lot of numbered items. And when I tried running our sentence splitter it turned strings like "1. Harry" into ["1.", "Harry"] consistently. And since these strings are now treated as 2 separate sentences, NLLB consistently translates "1." into "Ασφάλεια". Now why is "1." translated into "Ασφάλεια" I have no idea--but in general I found NLLB to have rather inconsistent outputs when the string is very short or in this case an artifact.

Hello,

Thanks for the fast response, glad you were able to identify the issue. Hope this can be fixed in a future version of the dataset. Let me know if I can do anything to help :)

gpantaz changed discussion status to closed

Sign up or log in to comment