This dataset is a flexible and lightweight drop in instruct dataset which is easisly scalable from 1-15K samples, intended to fill out the generalist assistant capabilites of specialized finetunes and subject matter experts.
The main feature of EzInstruct is that the samples are ordered as they were sampled by the NovelSelect sampling algorithm from the full Magpie-Align/Magpie-Llama-3.1-Pro-MT-300K-Filtered dataset.
This means that simply pulling the first 2000 samples with… See the full description on the dataset page:
https://huggingface.co/datasets/ConicCat/EzInstruct1-15K-MagpieProMT.
