KAIST reports up to 69.6 percent less fine-tuning time
KAIST announced CluSTER on November 2, after a September preprint and an October conference presentation. The method picks less-repetitive training data and divides it among GPUs. The team tested models with about 7 billion and 13 billion parameters on three coding and medical datasets, using eight performance benchmarks.
KAIST reports nearly the same performance with about 31 percent of the training data, cutting time by up to 69.6 percent. These are results from interested researchers on selected workloads, not a measured saving for every model or proof of medical benefit.
Read article →