YKS-LLM-v3
YKS-LLM-v3 is a text-only Large Language Model fine-tuned for Turkish exam-style mathematics questions (YKS).
Unlike earlier vision-based versions, this model operates purely on structured text input and focuses on reasoning quality, solution consistency, and exam-aligned outputs.
It is designed to be the core reasoning engine in the YKS pipeline, typically consuming cleaned / parsed text produced by an upstream OCR or VLM component.
check us out: cga-labs.com
⸻
What this model is good at • Solving Turkish YKS-style math questions from clean text • Producing step-by-step, human-readable solutions • Maintaining ÖSYM-like reasoning flow and notation • Handling multi-step algebra, geometry, and function problems • Generating stable, structured outputs for evaluation and downstream use • Acting as the final solver in a modular YKS system
This model is not responsible for: • OCR • Image understanding • Layout parsing
It assumes the question is already correctly extracted into text.
⸻
Typical output format
{ "answer": "C", "solution_steps": [ "Fonksiyonun tek olduğu bilgisi kullanılarak f(-x) = -f(x) yazılır.", "g fonksiyonunun çift olması nedeniyle g(-x) = g(x) elde edilir.", "(f + g)(-x) - (f + g)(x) ifadesi sadeleştirilir.", "Elde edilen polinom katsayıları karşılaştırılarak a ve b bulunur.", "Bulunan değerler f(2a + b) ifadesinde yerine yazılır ve sonuç hesaplanır." ] }
⸻
Design philosophy • Text-first reasoning: no visual noise, no hallucinated diagrams • Exam realism over flashy chain-of-thought • Structured outputs over free-form explanations • Composable architecture: works cleanly with VLMs, OCR pipelines, and evaluators
- Downloads last month
- 12