The paper proposes a data selection strategy to remove low-quality instances from instruction-finetuning datasets for large language models. The resulting model, AlpaGasus, outperforms the original model and achieves faster training time.

https://arxiv.org/abs//2307.08701

YouTube: https://www.youtube.com/@ArxivPapers

PODCASTS:

Apple Podcasts: https://podcasts.apple.com/us/podcast/arxiv-papers/id1692476016

Spotify: https://podcasters.spotify.com/pod/show/arxiv-papers