Reproduce “SODA: Million-scale Dialogue Distillation with Social Commonsense Contextualization” 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
Hyunwoo Kim, Jack Hessel, Liwei Jiang, Peter West, Ximing Lu, Youngjae Yu, Pei Zhou, Ronan Le Bras and 4 more
Researchers introduced SODA, the first publicly available million-scale dataset for open-domain social dialogue, distilled from large language models using a social commonsense contextualization framework (CO3). A model trained on SODA, COSMO, demonstrated superior generalizability and naturalness, with human raters preferring its responses over human-authored ground truth on an unseen dataset like DailyDialog.
Key insights
- Grounding large language model-based dialogue generation in social commonsense knowledge is crucial for distilling high-quality, diverse, and coherent conversations.
- The CO3 framework demonstrates that carefully contextualized and filtered machine-generated dialogues can achieve higher perceived quality by human raters than many existing human-authored datasets.
- Training smaller conversation models on such large-scale, high-quality distilled data enables superior generalizability and naturalness compared to larger, generically trained language models or previous state-of-the-art dialogue systems.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.