Mustafa Erşahin

Mustafa Erşahin

Auto-generated profile, based on publicly available information

Mustafa Erşahin

Research and Development Department, Commencis Teknoloji, Istanbul, Trkiye

Research Focus & Metrics

Academic Career Length

8Years (2019-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
15 May 2019 in Turkish Journal of Electrical Engineering and Computer Sciences
Verified
Journal Article
Closed access
BEBuket Erşahi̇n
Buket Erşahi̇n
ÖVÖzlem Varlıklar
Özlem Varlıklar
DKDeniz KILINÇ
Deniz KILINÇ
MEMustafa ERSAHIN
Mustafa ERSAHIN

31 May 2024 in Orclever Proceedings of Research and Development
Verified
Journal Article
Open access
PDF file
PINAR ERSOYPINAR ERSOY
PINAR ERSOY
MEMustafa ERSAHIN
Mustafa ERSAHIN
This study benchmarks the capabilities of Llama 3 70B, a 70-billion parameter large language model (LLM), for code generation tasks. To effectively train and fine-tune this massive model, we integrate PyTorch Fully Sharded Data Parallel (FSDP) [1], [2] for distributed training and Quantized Low-Rank Adaptation (Q-LoRA) [7] for efficient fine-tuning. We address challenges associated with distributed training, including communication overhead and synchronization complexities, through optimization strategies like gradient accumulation, optimizer state sharding, and mixed precision training. Additionally, we employ advanced training techniques such as Curriculum Learning, Dynamic Batch Sizing, and Adaptive Optimization Algorithms to enhance model performance and training efficiency. Our primary focus is evaluating the performance of the fine-tuned Llama 3 70B model on two widely-recognized code generation benchmarks: HumanEval [8] and MBPP [9]. HumanEval assesses the model's ability to translate natural language problem descriptions into functionally correct code, while MBPP evaluates its proficiency in solving complex programming problems by generating accurate Python code. We present detailed performance results on these benchmarks, analyzing the model's strengths and limitations in various code generation scenarios. Furthermore, we compare the impact of our training and fine-tuning methodologies on scalability, memory efficiency, and training speed, demonstrating the feasibility and efficiency of our approach. This benchmark study offers valuable insights for researchers and practitioners exploring the application of LLMs for code generation. It provides a comprehensive evaluation of Llama 3 70B's capabilities, sheds light on the effectiveness of various training and fine-tuning techniques, and emphasizes the importance of rigorous benchmark evaluation in driving progress within this rapidly evolving field.

13 November 2025 in IEEE Access
Verified
Journal Article
Open access
PINAR ERSOYPINAR ERSOY
PINAR ERSOY
MEMustafa ERSAHIN
Mustafa ERSAHIN
Retrieval-Augmented Generation (RAG) pairs large language models with external search to constrain knowledge staleness and hallucination, a critical need in finance and e-commerce where numerical precision and regulatory auditability are non-negotiable. This study resembles four paradigms: sparse, dense, hybrid, and fusion retrieval, each realized as LangChain templates for cloud-hosted GPT-4o and on-premises LLaMA deployments. We embed a zero-shot baseline and assess every pipeline with the Retrieval-Augmented Generation Assessment (RAGAS) protocol, which scores faithfulness, relevancy, context quality, and latency without revealing proprietary gold data. Hybrid and fusion systems achieve the best combination of recall and faithfulness. Dense retrieval maximizes semantic coverage; sparse retrieval provides clearer, faster traceability. Each yields substantial gains over the baseline by anchoring profit metrics, compliance evidence, and pricing to audited documents. Although empirical evaluation primarily focuses on finance and e-commerce, the same blueprints and metrics can apply to healthcare, education, and legal technology. The domains in scope demand reasoning that is traceable, bias-aware, and up to date. This study, therefore, provides an auditable, production-ready framework for reliable RAG across knowledge-intensive settings.

16 June 2025 in IEEE Access
Verified
Journal Article
Open access
Göksu TüysüzoğluGöksu Tüysüzoğlu
Göksu Tüysüzoğlu
Yunus DoğanYunus Doğan
Yunus Doğan
Elife Ozturk KiyakElife Ozturk Kiyak
Elife Ozturk Kiyak
MEMustafa ERSAHIN
Mustafa ERSAHIN
Bita GhasemkhaniBita Ghasemkhani
Bita Ghasemkhani
Derya BirantDerya Birant
Derya Birant
Noisy data is a prevalent issue in data mining, significantly impacting the performance of classification algorithms. Mathematical methods are crucial in tackling this obstacle, particularly in optimizing noise detection and data preprocessing. This study proposes a novel approach—Joint Tomek Links (JTL)— to identify and eliminate noisy instances by detecting pairs of nearest neighbors from different classes. It first finds the Tomek links and then refines a probabilistic method to determine which instance from a pair will be removed. In our approach, a random tree classifier serves as the base model. We conducted experiments on 40 benchmark datasets spanning various domains, achieving an average classification accuracy of 83.26% for JTL. The results demonstrate that the JTL attains an average improvement of 5.33% in accuracy compared to the original classification with a random tree. Furthermore, JTL surpasses existing techniques, delivering a noteworthy gain in accuracy by 12.30% on the same datasets. These findings underscore the effectiveness of JTL in enhancing data quality and boosting classification performance in data mining tasks.

23 November 2025 in The European Journal of Research and Development
Verified
Journal Article
Open access
PDF file
PINAR ERSOYPINAR ERSOY
PINAR ERSOY
MEMustafa ERSAHIN
Mustafa ERSAHIN
Aspect-based sentiment analysis provides granular insights into customer feedback by identifying discrete aspects, such as features or topics, and assigning a corresponding sentiment to each. This study assesses three large language models, hereafter referred to as LLMs, namely Google Gemini 2.5 Flash-Lite, Anthropic Claude Sonnet-4 delivered through AWS Bedrock, and Meta LLaMA 3.3 70B delivered through AWS Bedrock, using a real-world multilingual corpus of 7,841 Turkish mobile banking app reviews from İşbank in Turkey. We employ a prompt-based tagging protocol to extract aspect–sentiment pairs from every review, and we compare accuracy, F1-score, inference cost, and latency. The results show that all three LLMs can execute multilingual aspect extraction and sentiment categorization without task-specific fine-tuning. Claude Sonnet-4 attains the highest F1 for aspect extraction and the highest sentiment accuracy, although it incurs a markedly higher inference cost. Gemini 2.5 Flash-Lite achieves competitive accuracy at a fraction of the price, making it well-suited for high-volume analytics. Meta LLaMA at the 70B scale accessed through AWS Bedrock exhibits intermediate performance with moderate cost and latency. We provide detailed performance tables and figures, along with best-practice guidance for enterprise deployment. AWS Bedrock enables the strategic selection of Claude and LLaMA 3.3 70B for multilingual sentiment analysis, offering valuable insights from app reviews within scale, accuracy, and budget constraints.