taq_Tfng is contaminated with zgh_Tfng data
I've started a discussion here.
This explains why GlotLID misclassifies zgh as taq, even though it should be easy to tell them apart.
Thank you for the report. To be clear, we are talking about the Tfng script, right?
https://huggingface.co/datasets/cis-lmu/glotlid-corpus/tree/main/v3.1/taq_Tfng
https://huggingface.co/datasets/cis-lmu/glotlid-corpus/tree/main/v3.1/zgh_Tfng
Among these two links, which exact sources have contamination? FLORES, Tatoeba, or NLLB-seed?
Which one should be deleted? What are the alternative sources to add?
Thank you for the report. To be clear, we are talking about the Tfng script, right?
Yes, the latin subset of Tamasheq looks ok.
Among these two links, which exact sources have contamination? FLORES, Tatoeba, or NLLB-seed?
The only source I've identified so far is NLLB-SEED (zgh samples intaq, here). There could be others but it's difficult to determine since I can only see the text, not the source.
Which one should be deleted? What are the alternative sources to add?
If you'd like to fix the issue on your end, you could deduplicate taq based on zgh OLDI-seed data. 4000 out of the 6193 samples are indeed Tamasheq.
So, the zgh part of OLDI is correct? We should just remove those from taq.
So, the zgh part of OLDI is correct? We should just remove those from taq.
Indeed. You're using the FLORES dev set forzgh, which is fine.
Hi, I ran a quick method to remove taq_Tfng samples with over 65% overlap with zgh_Tfng from https://huggingface.co/datasets/cis-lmu/glotlid-corpus/blob/main/v3.1/taq_Tfng/taq_Tfng_nllbSeed.txt. Could you check if it's clean now? Thanks.
taq_Tfng_nllbSeed.txt: Removed 201, Kept 5996. Should I remove more? You mentioned that 4000 out of the 6193 samples are indeed Tamasheq.
There are still a lot of zgh_Tfng samples.
I suggest you use the original NLLB-SEED files (not the OLDI SEED dataset on HF) and keep the first 4000 rows. Everything beyond that is zgh_Tfng and should be removed.
If you'd like to double check on your side, search for this letter: ⵗ. You'll find it in the Tamasheq samples (first 4000) but not elsewhere, because it's not used to write Standard Moroccan Tamazight.
Done. Thank you so much.