From Production Traffic to Post-Training: Building a Self-Hosted LLM That Covers the Corporate Request Mix
A single self-hosted model took over half of an enterprise LLM platform’s traffic.
The paper says the team consolidated requests from more than 200 internal applications by targeting failures in instruction following, function calling, and internal task mix. Instead of one joint optimization run, they trained separate GRPO experts for each axis and merged them with two-stage SLERP. Their non-reasoning setup beat a roughly 7x larger baseline on in-house Arena, instruction-following, and function-calling scores. The resulting model handles 116M requests a month at lower serving cost. HF Daily Papers' note
The paper says the team consolidated requests from more than 200 internal applications by targeting failures in instruction following, function calling, and internal task mix. Instead of one joint optimization run, they trained separate GRPO experts for each axis and merged them with two-stage SLERP. Their non-reasoning setup beat a roughly 7x larger baseline on in-house Arena, instruction-following, and function-calling scores. The resulting model handles 116M requests a month at lower serving cost. HF Daily Papers' note
score 5