This paper introduces InvestLM, a large language model specifically designed for financial investment applications, built by instruction tuning the LLaMA-65B foundation model. The key innovation lies in the manual curation of a small but diverse instruction dataset, drawing from sources such as Chartered Financial Analyst (CFA) exam questions, SEC filings, and Stackexchange quantitative finance discussions. This approach is inspired by the 'less-is-more-for-alignment' principle and aligns with the Superficial Alignment Hypothesis, demonstrating that high-quality domain-specific models can be developed with minimal, carefully selected data rather than massive datasets.

InvestLM exhibits strong performance in understanding financial texts and providing investment-related advice, as validated through evaluations by financial experts including hedge fund managers and research analysts. These experts rated its responses as comparable to leading commercial models like GPT-3.5, GPT-4, and Claude-2. Additionally, zero-shot testing on financial NLP benchmarks confirms the model's generalizability. From a research perspective, this work contributes to the field by supporting efficient domain adaptation strategies for LLMs. Practically, it offers a state-of-the-art tool that could enhance the productivity of financial professionals, with the model parameters being released to the research community to encourage further exploration and application.