Enhancing Reasoning Skills in Small Persian Medical Language Models Can Outperform Large-Scale Data Training
Bibliographic record
Abstract
Reasoning is a critical requirement for medical language models, where incorrect or poorly justified outputs can have serious consequences. Despite the importance of this capability, prior work has largely focused on high-resource languages, leaving Persian—a widely spoken and medically relevant language—significantly underexplored. In this work, we address this gap by introducing two complementary post-training frameworks designed to enhance medical reasoning in Persian language models. Our central hypothesis is that, while training on massive medical corpora can improve factual knowledge, targeted post-training using preference-based optimization methods can yield greater gains in medical reasoning efficiency and reliability, even with substantially less data. The newly introduced model shares the same baseline as our previous model, gaokerena-V, which was obtained by training the aya-expanse-8b on 57 million tokens of web-crawled Persian medical data, in this work we apply small-scale post-training using Direct Preference Optimization (DPO) and Reinforcement Learning from AI Feedback (RLAIF). This post-training relies on only 11,000 AI-generated preferred–rejected response pairs, comprising approximately 2 million tokens in preferred answers and 2.5 million tokens in rejected ones, without requiring costly human annotation. Experimental results on a Persian-translated Medical MMLU benchmark demonstrate that this lightweight post-training approach outperforms gaokerena-V by a margin of 3.67 percent, highlighting that structured preference-based reasoning supervision can be more effective than large-scale domain data training alone for improving medical reasoning in low-resource languages.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.003 | 0.010 |
| Meta-epidemiology (narrow) | 0.001 | 0.000 |
| Meta-epidemiology (broad) | 0.001 | 0.001 |
| Bibliometrics | 0.000 | 0.000 |
| Science and technology studies | 0.000 | 0.001 |
| Scholarly communication | 0.001 | 0.002 |
| Open science | 0.001 | 0.001 |
| Research integrity | 0.001 | 0.003 |
| Insufficient payload (model declined to judge) | 0.006 | 0.003 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".