Wnuan: A Three-Stage Pipeline for Enterprise QA with Proprietary Knowledge

A recent study on arXiv introduced Wnuan, an innovative three-phase approach aimed at improving question answering within proprietary enterprise knowledge while maintaining general capabilities. The framework generates specific guidance from existing documents, employs supervised fine-tuning with general data replay, and integrates reinforcement learning to address remaining errors. In a benchmark test involving 707 questions, the primary 32B model enhanced the acceptable-answer rate from 52.76% to 80.06% after fine-tuning, and to 91.51% with reinforcement learning. Furthermore, a 100-update protocol indicated that residual-error sampling surpassed both full-pool and random sampling by 3.11 and 2.97 points, respectively.