Reproduce “SaFeRDialogues: Taking Feedback Gracefully after Conversational Safety Failures” 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
Megan Ung, Jing Xu, Y-Lan Boureau
SaFeRDialogues trains dialogue models to respond receptively when a conversation partner signals that a remark was unsafe. It pairs unsafe exchanges with human-written feedback and civil recovery responses, while mixing this data with ordinary chat during fine-tuning. Human raters preferred Recovery (DialoGPT) over a BST-fine-tuned baseline 86% to 14% on feedback contexts; this measures judgments, not future behavior or safety.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.