‫ﺗﻮﺳﻌﻪ ‬‫ﯾﮏ ‬‫ﻣﺪل‬ ‫زﺑﺎﻧﯽ‬ ‫ﭘ‬ﺰﺷﮑﯽ‬ ‫ﻣﺒﺘﻨﯽ‬ ‫ﺑﺮ‬ ‫اﺳﺘﺪﻻل در زبان فارسي

The use of artificial intelligence in answering medical questions has attracted significant attention in recent years as an emerging an‎d important area in technology an‎d healthcare. While extensive research in this field has been published in English, at the time this thesis began, the Persian language domain was lacking in research. In the first phase of this study, a medical corpus consisting of ninety million tokens an‎d a dataset comprising twenty thousan‎d medical questions an‎d answers provided by physicians was collected from journals an‎d online QA forums. Using part of this data, the base model "Aya-Expanse" was trained, an‎d a new model named "Gaokerena-V" was introduced, which achieved 2.67% better performance compared to its base model on the MMLU medical dataset translated into Persian; in the second phase, considering that medicine is heavily reliant on reasoning an‎d logical analysis, a new model based on a chain of thoughts an‎d logical reasoning was designed to significantly enhance its accuracy an‎d efficiency. For this purpose, two new frameworks based on reinforcement learning with AI feedback were introduced to improve the reasoning capabilities of the base model, Aya-Expanse, through direct optimization of preferences. The result of this improvement was the "Gaokerena-R" model, which, despite being trained on a significantly smaller number of tokens compared to Gaokerena-V, achieved 6.34% better performance than its base model on the same MMLU medical dataset translated into Persian. This level of improvement highlights the importance of reasoning in the field of medicine.

7

167681