From Production Traffic to Post-Training: Building a Self-Hosted LLM That Covers the Corporate Request Mix

Hugging Face Daily Papers Papers

Summary

A research paper presents a method for training a smaller self-hosted LLM using separate GRPO experts merged via SLERP, which outperforms a larger baseline on enterprise tasks and serves half of platform traffic at lower cost.

Data-residency constraints force enterprises to self-host LLMs, but continuous adoption of newer models without decommissioning their predecessors expands the serving fleet, fragmenting a finite GPU pool. We consolidate traffic from over 200 internal applications onto a single model by closing quality gaps identified through production error analysis along three axes: instruction following, function-calling, and internal task distribution. Quality is tracked by offline benchmarks stratified to production traffic and scored by deterministic verifiers or calibrated LLM judges. Rather than optimising all objectives jointly, which introduces cross-domain reward interference, we train a separate GRPO expert per axis and merge them via two-stage SLERP. Each expert's reward exposes a distinct failure mode, namely semantic collapse, over-calling, and verbosity hacking, each requiring a domain-specific fix. In non-reasoning mode the recipe surpasses a {sim}7times larger by total parameters baseline on the in-house Arena with 69.6 to 65.8, instruction following with 0.85 to 0.83, and function-calling with 0.79 to 0.77, while lifting general dialogue benchmarks. The model absorbs 50% of platform traffic, 116M requests per month, at a fraction of the serving cost.
Original Article
View Cached Full Text

Cached at: 09/02/26, 11:45 AM

Paper page - From Production Traffic to Post-Training: Building a Self-Hosted LLM That Covers the Corporate Request Mix

Source: https://huggingface.co/papers/2609.01572 Authors:

,

,

,

,

,

,

,

,

,

,

,

,

Abstract

A smaller self-hosted LLM trained with separate GRPO experts merged via SLERP outperforms a much larger baseline on instruction following, function-calling, and internal tasks while serving half of platform traffic at lower cost.

Data-residency constraints force enterprises to self-host LLMs, but continuous adoption of newer models without decommissioning their predecessors expands the serving fleet, fragmenting a finite GPU pool. We consolidate traffic from over 200 internal applications onto a single model by closing quality gaps identified through production error analysis along three axes:instruction following,function-calling, and internal task distribution. Quality is tracked by offline benchmarks stratified to production traffic and scored bydeterministic verifiersorcalibrated LLM judges. Rather than optimising all objectives jointly, which introduces cross-domainreward interference, we train a separateGRPOexpert per axis and merge them via two-stageSLERP. Each expert’s reward exposes a distinct failure mode, namelysemantic collapse,over-calling, andverbosity hacking, each requiring a domain-specific fix. In non-reasoning mode the recipe surpasses a {sim}7times larger by total parameters baseline on the in-house Arena with 69.6 to 65.8,instruction followingwith 0.85 to 0.83, andfunction-callingwith 0.79 to 0.77, while lifting general dialogue benchmarks. The model absorbs 50% of platform traffic, 116M requests per month, at a fraction of the serving cost.

View arXiv pageView PDFProject pageAdd to collection

Get this paper in your agent:

hf papers read 2609\.01572

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2609.01572 in a model README.md to link it from this page.

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2609.01572 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2609.01572 in a Space README.md to link it from this page.

Collections including this paper0

No Collection including this paper

Add this paper to acollectionto link it from this page.

Similar Articles