Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).
We present Yway (ရွေး, "choose"), a small decision model for Burmese (Myanmar) text. Instead of generatingtext, Yway answers typed questions (pick an option, yes/no, or a level) with calibrated probabilities, following the"System One" design of TypeSafe's Jev and the open Laya model. It reuses Laya's decision head and itsreinforcement-learning objective with proper scoring rules, but replaces Laya's multilingual encoder with XLMRoBERTa-base, which tokenizes Burmese 2.3× more efficiently. All training data is derived from BurmeseWikipedia: 37,159 encyclopedic articles, 251,642 question–answer pairs generated with Gemma 3 27B, and topiclabels from Wikipedia categories. On held-out articles with question wordings never seen in training, Yway routesquestions by type with 99.4% accuracy, decides whether a passage answers a question with 91.6%, and assigns oneof 12 topics with 85.8%, with low calibration error on most tasks. An answer checker fine-tuned on 22,000 realanswers of Gemma 3 12B, graded by Gemma 3 27B, catches 81% of wrong answers at a 0.5 threshold, raising thequality of shown answers from 0.44 to 0.68. Yway does not answer genuinely new kinds of questions zero-shot,but learns a new decision from about 300 labelled examples in minutes without forgetting. The model is served ona CPU through a Jev-compatible API.
No comments yet — start the discussion below.