|
|
--- |
|
|
license: cc-by-sa-4.0 |
|
|
--- |
|
|
Performance: 11.27 BLEU, 52.25 ChrF |
|
|
|
|
|
PumaNMT is a parallel dataset composed of 8,000 sentences in Nepali and Puma. It is intended to be used for fine-tuning Neural Machine Translation models and Large Language Models for Puma. |
|
|
|
|
|
Puma is a low-resource Sino-Tibetan language spoken in Nepal and nearby regions. |
|
|
|
|
|
This dataset was created and curated using a proprietary method developed by XRI Global which ensures coverage of a conceptual space when doing data collection. This method was developed in order to create the fastest and most affordable way to collect pristine in-domain data for low-resource languages optimized for fine-tuning language models. |
|
|
|
|
|
The contributors gave proper consent, were hired by a local agency, and were compensated fairly. |
|
|
|
|
|
For data collection, we utilized our mobile data collection app, Echonet, and our custom translation management system. |
|
|
|
|
|
The domain is somewhat generalized and the dataset is most effective at literary and narrative texts. It is less capable in other domains such as technical, scientific, or colloquial. |
|
|
|
|
|
If you are interested in creating ideal datasets for new languages, reach out to us at contact@xriglobal.ai |
