Reproduce “OpenCharacter: Training Customizable Role-Playing LLMs with Large-Scale Synthetic Personas” 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
Xiaoyang Wang, Hongming Zhang, Tao Ge, Wenhao Yu, Dian Yu, Dong Yu
Researchers at Tencent AI Lab Seattle developed OpenCharacter, a method to train Large Language Models for customizable, out-of-domain role-playing using large-scale synthetic personas and character-driven dialogue synthesis. Their 8B OpenCharacter model achieves performance on persona generalization benchmarks comparable to or exceeding larger commercial models like GPT-3.5 and earlier GPT-4o versions.
Key insights
- A strong instruction-following backbone model (e.g., LLaMA-3-8B-Instruct) is a critical prerequisite for effective character generalization, outperforming base models.
- Direct response generation (OpenCharacter-G) for synthetic data is more effective than rewriting existing responses (OpenCharacter-R), leading to better model performance.
- LLaMA-3-70B-Instruct can serve as an effective prompting model for synthesizing high-quality training data, sometimes outperforming GPT-4o for this task.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.