This paper explores the adaptation of large language models (LLMs) for financial sentiment analysis, focusing on smaller models with up to 1.5 billion parameters. The authors adapt two foundation models—OPT and Pythia—using various strategies, including further pre-training on financial documents and fine-tuning on instructions. They demonstrate that these smaller models can achieve comparable or better performance compared to larger models, while being more efficient in terms of parameters and data usage. The study also introduces a method for generating synthetic instructions to augment the instruction dataset, enhancing the diversity and richness of the training data. The results show that the proposed domain adaptation methods are effective for multiple financial NLP tasks, outperforming state-of-the-art models in classification tasks such as financial sentiment analysis and named entity recognition. The paper concludes by discussing the limitations and future directions, including the potential of applying similar fine-tuning strategies to larger models and exploring methods like Low-Rank Adapters (LoRA) for further improvements.This paper explores the adaptation of large language models (LLMs) for financial sentiment analysis, focusing on smaller models with up to 1.5 billion parameters. The authors adapt two foundation models—OPT and Pythia—using various strategies, including further pre-training on financial documents and fine-tuning on instructions. They demonstrate that these smaller models can achieve comparable or better performance compared to larger models, while being more efficient in terms of parameters and data usage. The study also introduces a method for generating synthetic instructions to augment the instruction dataset, enhancing the diversity and richness of the training data. The results show that the proposed domain adaptation methods are effective for multiple financial NLP tasks, outperforming state-of-the-art models in classification tasks such as financial sentiment analysis and named entity recognition. The paper concludes by discussing the limitations and future directions, including the potential of applying similar fine-tuning strategies to larger models and exploring methods like Low-Rank Adapters (LoRA) for further improvements.