Incentivizing Inclusive Data Contributions in Personalized Federated Learning

The preprocessed datasets used in our experiments are provided in the data folder. For the image or character classification task, we use 5 classical datasets: Cifar-10, Fashion-MNIST, PACS, FEMNIST, and Shakespeare. We consider the mixed-finance and code+finance scenarios for instruction-tuning tasks, involving 3 financial datasets (TFNS, FIQA, NWGI) and a code dataset (CodeAlpaca). CIFAR-10 and Fashion-MNIST are widely used benchmarks in literature for image classification tasks containing 10 categories. PACS has four domains (photo, art painting, cartoon, and sketch) and contains seven categories. FEMNIST for image classification and Shakespeare for the next character prediction are from the naturally heterogeneous synthetic dataset Leaf. Three finance datasets include: FiQA comprised of 17k sentences sourced from microblog headlines and financial news, The Twitter Financial News Sentiment (TFNS) with 11,932 annotated documents of finance-related tweets, and the News With GPT Instruction (NWGI)featuring labels generated by ChatGPT. The code dataset CodeAlpaca contains 20K instruction-following data. Note that all raw data resources can be found in the "Data availability" section in our paper.