Reproduce “Improving LLM Safety and Helpfulness using SFT and DPO: A Study on OPT-350M” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Piyush Pant
The authors test whether sequential supervised fine-tuning and direct preference optimization improve alignment on OPT-350M. They train on preferred answers, then preferred-versus-rejected response pairs, and evaluate four variants on 100 selected prompts using a reward model. The sequence scores highest in helpfulness (66%) and combined score (55%), but SFT has higher harmlessness (48% versus 44%).
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.