From Production Traffic to Post-Training: Building a Self-Hosted LLM That Covers the Corporate Request Mix
TL;DR - This paper presents a post-training recipe for consolidating traffic from more than 200 enterprise applications onto one self-hosted LLM. Specialized GRPO experts and model merging close production-specific quality gaps while reducing GPU fragmentation and serving costs.
- Production errors are grouped into instruction following, function calling, and internal task-distribution gaps, with evaluations stratified to real traffic.
- Separate GRPO experts avoid cross-domain reward interference and are combined using two-stage SLERP.
- Domain-specific fixes address semantic collapse, excessive function calls, and reward-driven verbosity.
- The resulting model reportedly beats a roughly 7× larger baseline on internal Arena, instruction-following, and function-calling evaluations, and serves 116 million monthly requests—50% of platform traffic.