Question about local client data size

[mynameisissac]

Hi, may I ask how you decide the number of training samples per client? and what is the exact number used, for example, in arithmetric reasoning experiment using MetaMathQA?

I couldn't find numbers about local data size in the paper but I think they are important details in the federated setup.

[josueCastillo0272]

Hi! As configured in fed_sb/fed/scripts/arithmetic.sh, the experiment utilizes the first 20,000 training samples from the MetaMathQA dataset.

These samples are partitioned evenly among the default 25 clients, resulting in a local data size of 800 samples per client.