1 results about "Range constraint" patented technology

The application provides a language model federated fine-tuning method and device and electronic equipment. The method comprises: receiving first information sent by N clients, wherein the first information comprises a data importance index of the client; determining ideal ranks of the N clients based on the data importance index and a preset total rank budget; determining assigned ranks of the N clients that satisfy a preset constraint condition according to the ideal ranks of the N clients and a preset delay; the preset constraint condition comprises at least one constraint: an efficiency constraint, a rank budget constraint and a rank range constraint; sending corresponding assigned ranks to the N clients respectively, wherein the assigned ranks are used to train local low-rank adaptive matrices of the clients; receiving trained local low-rank adaptive matrices sent by the N clients; and adjusting a weight matrix of the language model based on the trained local low-rank adaptive matrices of the N clients. The performance of the language model can be improved.