Reimagining Supervised Fine-Tuning for Large Language Models and Diffusion from a Denoising Perspective

TL; DR 🚧 (WIP) In a recent study by Gekhman, Zorik, et al., the authors reveal that supervised fine-tuning (SFT) struggles to integrate new knowledge into Large Language Model (LLM). The authors propose an approach by categorizing fine-tuning datasets into four categories as below: Their findings suggest that the most effective fine-tuning datasets should consist of HighlyKnown, MaybeKnown, and WeaklyKnown data, with minimal amount of Unknown data. This conclusion makes me rethink the natur...