Scaling Instruction Fine-Tuning in LLMs: A Comparative Review of Datasets, Pipelines, and Open-Source Implementations

This review paper presents a comprehensive and structured analysis of instruction fine-tuning in large language models (LLMs). It synthesizes recent developments in datasets, supervised fine-tuning (SFT), alignment techniques such as RLHF and DPO, and widely used evaluation frameworks including MT-Bench, HELM, and GPT-4-based assessments. The work compares leading open-source instruction-tuned models, highlights scaling trends, and provides a taxonomy of tools and pipelines used in modern LLM development. This preprint aims to support researchers and practitioners by offering a unified understanding of the key components, challenges, and future directions in instruction fine-tuning.