BAAI Proposes Wnuan: 3-Stage Post-Training for Enterprise QA

BorisMPower · x · 2026-08-05

BAAI introduced Wnuan, a post-training pipeline for question answering over proprietary enterprise knowledge, aiming to acquire enterprise knowledge without discarding general capabilities.

Methodology

Wnuan utilizes a three-stage pipeline:

Constructs task-oriented supervision from documents.
Performs supervised fine-tuning (SFT) with general-data replay.
Applies reinforcement learning (RL) to residual errors.

Results

On the 707-question WnuanBench, the 32B route raised the acceptable-answer rate from 52.76% (pre-adaptation) to 80.06% (post-SFT), reaching 91.51% after RL.
Under a matched 100-update protocol, residual-error sampling outperformed full-pool and size-matched random sampling.
The trade-off is a 5.17-point decrease in the general-benchmark average, primarily concentrated in instruction following.

Original post →

BAAI Proposes Wnuan: 3-Stage Post-Training for Enterprise QA

More from Research