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

2026-07-02 · Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering · Depth: Expert, quick

Summary

TUDUM (Türkçe Düşünen Üretken Model) is a project pipeline designed to adapt the Qwen3.5-27B thinking model for explicit Turkish reasoning, ensuring the internal reasoning trace is Turkish, not just the final output. The pipeline begins with the unsloth/Qwen3.5-27B base checkpoint, applying supervised fine-tuning (SFT) on 15,991 Turkish reasoning examples using LoRA adapters. This SFT phase resulted in shorter, more consistently Turkish reasoning behavior and reduced thinking exhaustion, though it decreased benchmark accuracy. Subsequently, GRPO-family reinforcement learning was applied within a proxy-filtered Turkish mathematics environment. While RL recovered some mathematical performance, notably on AIME24 at an early checkpoint, it did not uniformly improve all benchmarks and failed to surpass the base model's Macro-6 average. The project is presented as an honest evaluation of a Turkish-thinking reasoning pipeline, with a step-50 model publicly available.

Key takeaway

For NLP Engineers developing localized LLMs, this research highlights the complexity of achieving truly native reasoning. If your goal is deep linguistic adaptation beyond surface translation, you should anticipate that supervised fine-tuning may localize reasoning but reduce benchmark accuracy. Consider a multi-stage approach, using reinforcement learning to recover specific task performance after initial SFT, and be prepared to evaluate "thinking" quality separately from final answer accuracy.

Key insights

Adapting LLMs for deep linguistic reasoning requires explicit fine-tuning of internal thought processes, not just output translation.

Principles

• SFT can localize reasoning but may reduce accuracy.
• RL can recover specific task performance.
• Explicitly training reasoning traces is distinct from output localization.

Method

The pipeline involves SFT on 15,991 Turkish reasoning examples using LoRA, followed by GRPO-family reinforcement learning in a proxy-filtered Turkish mathematics environment.

In practice

• Consider SFT for language-specific reasoning localization.
• Evaluate RL for targeted performance recovery post-SFT.
• Release intermediate models for community evaluation.

Topics

• Turkish NLP
• Large Language Models
• Reasoning Pipelines
• Supervised Fine-tuning
• Reinforcement Learning
• Qwen3.5-27B

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• LANG: Reinforcement Learning for Multilingual Reasoning with Language-Adaptive Hint Guidance — LANG enhances multilingual LLM reasoning by using language-adaptive hints with a progressive decay schedule, preventing language drift.
• Reasoning Quality Emerges Early: Data Curation for Reasoning Models — Reasoning quality in LLMs can be effectively improved by curating SFT data based on early token loss patterns, significantly boosting efficiency.
• Deep Dive into LLMs like ChatGPT — LLMs are trained in stages, from vast internet data to human-guided refinement, enabling complex, yet fallible, capabilities.
• Disentangling Linguistic Relatedness from Task Alignment in Cross-Lingual Transfer — Cross-lingual fine-tuning primarily improves task-format alignment, not language-specific knowledge transfer, especially for weaker models.
• Learning to Reason by Analogy via Retrieval-Augmented Reinforcement Fine-Tuning — RA-RFT teaches LMs analogical reasoning by training a retriever on reasoning benefit and fine-tuning with retrieved demonstrations.
• Expert Masking for More Efficient Expert Offloading — Fine-tuning can degrade existing factual knowledge, but strategies like parameter restriction or self-distillation can mitigate this.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.