Closed-source LLMs like ChatGPT have gained prominence for natural language tasks, including Entity Matching (EM), without requiring task-specific architectures. However, open-source LLMs, including Meta’s LLaMA series and models such as SOLAR and Yi, offer viable alternatives, albeit often requiring fine-tuning for optimal performance. This paper explores Domain Instruction Tuning (DIT), a fine-tuning method using domain-specific data, for EM. We evaluate models from 7B to 34B parameters on an EM instruction-tuning benchmark, introducing a novel data reduction pipeline that enhances training efficiency without compromising F1 scores. Our results show that Yi-34b, with 122.8M trainable parameters, surpasses pre-trained models including Ditto and HierGAT, achieving an F1 score 6.9 points higher than GPT-4-0613 across seven tasks and outperforming all GPT-3 class models on two of three unseen benchmarks in zero-shot settings.