AI4Fin: Fine-Tuning AI for Egyptian Financial Sentiment and Slang

Although the domain of financial sentiment analysis has been dramatically altered by the rise of large language models, existing solutions have yet to be adapted to work on the Egyptian discourse around financial markets. The vast majority of financial language models, like FinBERT, are trained exclusively on English-language datasets, while general-purpose language models for Arabic can work on dialectal texts, they are not tuned for finance-related information and cannot handle the informal, heavily code-switched nature of the language used to discuss finances in Egypt. This paper presents AI4Fin, which trains an open-source large language model for fin-sentiment analysis, fin-summary generation and expression explanation of Egyptian financial discourse with the emphasis on dialectal, code-switched and slang languages. Datasets were prepared with texts scraped from social media posts, articles written in Formal Standard Arabic and artificially constructed datasets of financial slang. These datasets were processed in order to preserve dialectical features, labeled using weak label-ing through CamelBERT (especially for social-media scraped) followed by human post-processing, revision, and balanced without leakage. The LLaMA-3.1-8B model was fine-tuned using QLoRA, an approach for efficient parameter tuning in an eight-billion parameter model on a single consumer graphics card, accomplishing three tasks — classification,summarization and expression explanation — in one model,reaching accuracy 84.79 and macro-F1 of 0.77, and in a head-to-head comparison it outperformed GPT-5.5 by over thirty macro-F1 points. The fine-tuned model was deployed within a hybrid retrieval-augmented chatbot for end-user interaction.