Arkios: An Open Bilingual English-Nepali Language Model Trained From Scratch, with a Devanagari-Aware Tokenizer
The episode discusses 'Arkios: An Open Bilingual English-Nepali Language Model Trained From Scratch, with a Devanagari-Aware Tokenizer.' Hosts examine how this model was built from scratch using 150B tokens of bilingual text. They focus on the importance of script awareness via a custom tokenizer and discuss improvements like manifest-conditioned tool-use contracts for better reliability in low-resource language processing.