TÜDÜM: A Turkish-Thinking Reasoning Pipeline for Qwen3.5-27B

Sep 21, 2026

Uploaded by Violet

This paper presents TÜDÜM (Türkçe Dü̧sünen Üretken Model), a project pipeline adapting a Qwen-family 27B thinking model toward Turkish reasoning. It targets a key distinction: not only producing Turkish final answers, but also making the explicit reasoning trace inside ... Turkish. The pipeline applies supervised fine-tuning on 15,991 Turkish reasoning examples with LoRA adapters, then uses GRPO-family reinforcement learning on a proxy-filtered Turkish mathematics environment. Results are mixed: SFT shortens responses and increases Turkish consistency yet lowers benchmark accuracy; RL recovers some math performance without surpassing the base model’s Macro-6 average. A step-50 model is released publicly.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages6

Reading Time15 min

AccessFree to download

Abstract

I. Introduction

II. Method

A. Foundation Model and Thinking Format

TÜDÜM: A Turkish-Thinking Reasoning Pipeline for Qwen3.5-27B Baran Bingöl Department of Artificial Intelligence and Data Engineering Faculty of Engineering Ankara University Ankara, Türkiye [22290007@ogrenci.ankara.edu.tr](22290007@ogrenci.ankara.edu.tr) Bahaeddin Türko˘glu Department of Artificial Intelligence and Data Engineering Faculty of Engineering Ankara University Ankara, Türkiye [turkoglub@ankara.edu.tr](turkoglub@ankara.edu.tr) arXiv :2607 .0 1927v 1 [ cs .CL] 2 Jul 2026 Abstract—This paper presents TÜDÜM (Türkçe Dü¸sünen Üretken Model), a project pipeline for adapting a Qwen-family 27B thinking model toward Turkish reasoning. The central problem is not only to answer Turkish prompts in Turkish, but to make the explicit reasoning trace itself Turkish. A thinking model may translate a Turkish prompt into an English-centered internal or visible scratchpad, solve the problem mostly in English, and only localize the final answer. TÜDÜM instead treats the generated ... block as a trainable behavior. The pipeline starts from the project base checkpoint unsloth/Qwen3.5-27B, applies supervised fine-tuning (SFT) on 15,991 Turkish reasoning examples using LoRA adapters, and then applies GRPO-family reinforcement learning on a proxyfiltered Turkish mathematics environment. The results are mixed. SFT made the model shorter and more consistently Turkish in its reasoning behavior, with large reductions in average response length and thinking exhaustion, but reduced benchmark accuracy. RL recovered some mathematical performance, especially AIME24 at the best early checkpoint, yet did not uniformly improve all benchmarks and did not exceed the base model on the reported Macro-6 average. The contribution is therefore best framed as a technically honest Turkish-thinking reasoning pipeline and evaluation, not as a claim of state-of-the-art Turkish reasoning. The released step-50 model is publicly available. Model availability: [https://huggingface.co/barandinho/qwen3.5-27b-tudum-dapo-50](https://huggingface.co/barandinho/qwen3.5-27b-tudum-dapo-50) Index Terms—Turkish reasoning, multilingual reasoning, supervised fine-tuning, reinforcement learning, GRPO I. INTRODUCTION Reasoning models expose intermediate text before the final answer. In the TÜDÜM project this behavior is represented asa thinking block followed by a final response: intermediate reasoning trace final answer This output format changes the multilingual problem. A model can receive a Turkish prompt and produce a Turkish final answer while still using English for most of the reasoning trace. This behavior may be acceptable for making the answer fit the local language. But this is not reasoning in Turkish. This difference is important for education, for checking how the answer was produced, and for later use in Turkish-language workflows. Because the user or evaluator does not only see the final answer. They also see the order of the mathematical, scientific, or coding decisions that led to that answer. The project base checkpoint is unsloth/Qwen3 .5-27B. Its model card lists Qwen/Qwen3 .5-27B as the base model, and the official Qwen3.5 model card states that Qwen3.5 models operate in thinking mode by default [1], [2] . For broader Qwen-family background, we also cite the Qwen3 technical report [3]. Prior work on chain-of-thought prompting showed that explicit intermediate steps can improve complex reasoning in large language models [4] . Multilingual chain-ofthought work further showed that reasoning can transfer across languages at sufficient scale [5] . However, newer analyses also suggest that multilingual models can use English-centered internal representations even when both input and output are non-English [6] . TÜDÜM is motivated by this gap: Turkish output is not sufficient evidence that the model reasons in Turkish. This paper summarizes the TÜDÜM pipeline and its measured behavior on th

FAQs

What overall performance trend is reported for SFT and RL?

SFT makes reasoning shorter and more consistently Turkish but reduces benchmark accuracy; RL recovers some mathematical performance yet does not uniformly improve all benchmarks or exceed the base model on the reported Macro-6 average.

Sep 21, 2026

Uploaded by Violet

TÜDÜM: A Turkish-Thinking Reasoning Pipeline for Qwen3.5-27B_Page 1

Open in WPS to view all 6 pages