From Production Traffic to Post-Training: Building a Self-Hosted LLM That Covers the Corporate Request Mix
Olga Tsymboi, Dmitrii Stoianov, Ramil Latypov, Danil Taranets, Daniil Dryabin, Mikhail Gashkov, Viktor Zelenkovskiy, Aleksandr Fida, et al.
cs.CL
Quick verdict
Enterprise recipe: per-axis GRPO experts + SLERP merge rival a 7x larger LLM on in-house Russian benchmarks; strong ablations, self-judged headline metrics.
Summary
The paper describes how T-Tech consolidated traffic from 200+ internal applications onto a single self-hosted model built on Qwen3-32B, motivated by Russian data-residency constraints and GPU-pool fragmentation. Error analysis of 2,500 sampled production responses identified three weak axes: instruction following (37.9% of failures), function calling (~12% of traffic), and alignment to the internal task distribution. Because joint multi-objective GRPO showed cross-domain reward interference, the authors instead train three separate GRPO experts from a shared SFT checkpoint — general chat (reward model with a multiplicative length penalty and raised KL), instruction following (AutoIF-style synthetic Russian data; a verifier reward plus a reward-model penalty against semantic collapse), and function calling (1.2M English and 300K Russian natively generated dialogues; a binary Tool-N1 reward with synthetic irrelevance cases against over-calling) — and merge them via two-stage sequential SLERP. The merged non-reasoning model serves 116M requests per month (50% of platform traffic) and matches or surpasses the ~7× larger Qwen3-235B-A22B-Instruct-2507 on in-house Arena (69.57 vs 65.83), in-house IF (0.85 vs 0.83), and in-house BFCL (0.79 vs 0.77), at 2.8-3.9× lower per-token cost, while trailing on long-context, English, and multi-step retrieval tasks. An open-weight checkpoint (T-pro-it-2.1) trained without internal data shows similar public-benchmark gains. Key limitations: headline metrics rest on self-built, LLM-judged in-house benchmarks; only Russian and English are covered; single organization and model family.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Joint multi-domain GRPO over general/IF/FC rewards suffers cross-domain interference; training one GRPO expert per axis from a shared SFT checkpoint and merging via two-stage SLERP outperforms it at 32B (BFCLv3-Ru 65.96 vs 60.88; ruIFEval 0.799 vs 0.770; Table 8).
- The merged Qwen3-32B non-reasoning model surpasses the ~7x larger Qwen3-235B-A22B-Instruct-2507 on in-house Arena (69.57 vs 65.83), in-house IF (0.85 vs 0.83), and in-house BFCL (0.79 vs 0.77).
- Each expert's reward induces a distinct failure mode — semantic collapse (IF), over-calling (FC), verbosity hacking (general) — each fixed by a domain-specific reward or data-mixture correction rather than joint tuning.
- Template-aware traffic sampling plus task-specific judge routing reaches kappa 0.88 (reference-based) and 0.72 (open-ended) human agreement versus 0.63/0.57 for uniform side-by-side judging.
- A public checkpoint trained with the same recipe without internal data approaches the deployed model's public-benchmark scores, attributing the gains to the recipe rather than proprietary data.
Strengths
Rare production-scale validation
116M requests/month (50% of platform traffic) from 200+ services six months after rollout; concrete serving numbers (45 avg/110 peak RPS, single-GPU FP8 vLLM, p95 3.2s, TTFT 0.3s) and cost reductions of 2.8-3.9x per token, up to 4-9x for prior largest-model users (Section 5).
Benchmark construction rigor
Template-aware sampler tops diversity (Dist 0.953) while keeping JS distance to production low versus diversity-first samplers (Table 1); task-routed judging lifts human agreement kappa 0.63->0.88 reference-based and 0.57->0.72 open-ended (Table 2), with classifier accuracy 90.6-99.6% validated against annotator consensus (Table 10).
Unusually complete ablations
Actionable reward-hacking taxonomy
Recipe-not-data claim supported
Weaknesses
Headline wins are self-referential and judge-sensitive
Selective abstract framing
No uncertainty on headline deltas
Teacher-baseline coupling
Reproduction limited to recipe
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This is an industrial empirical study from T-Tech describing the consolidation of 200+ internal applications onto one self-hosted, Russian-focused LLM (Qwen3-32B with a Cyrillic-dense tokenizer, non-reasoning mode only) via production-driven post-training. Three claims anchor the paper: (1) joint multi-domain GRPO suffers cross-domain reward interference, so per-axis GRPO experts merged by two-stage SLERP is the better recipe (Tables 5, 7, 8, 23); (2) the merged model matches or beats the ~7x larger Qwen3-235B-A22B-Instruct-2507 on in-house Arena (69.57 vs 65.83), in-house IF (0.85 vs 0.83), and in-house BFCL (0.79 vs 0.77) while serving 116M monthly requests at 2.8-3.9x lower per-token cost; (3) each reward exposes a distinct failure mode — semantic collapse, over-calling, verbosity hacking — each fixed domain-specifically. All three are supported within their stated scope, which the Limitations section correctly bounds to Russian/English, one organization, one model family.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
Is T-pro-it-2.1 open source and where can I get it?
The paper releases an open-weight checkpoint, T-pro-it-2.1, on Hugging Face (t-tech/T-pro-it-2.1). It was trained with the same recipe as the deployed internal model but without any internal corporate data, and it scores close to the deployed version on public benchmarks (e.g., ruIFEval 0.807 vs 0.799), supporting the claim that the recipe rather than proprietary data drives the gains.
How can a 32B model beat a 235B model?
Only on specific metrics and in non-reasoning mode: the Qwen3-32B-based model beats Qwen3-235B-A22B-Instruct-2507 (a mixture-of-experts with ~22B active parameters, ~7x larger by total parameters) on the company's in-house Arena (69.57 vs 65.83), in-house instruction-following (0.85 vs 0.83), and in-house BFCL (0.79 vs 0.77) after targeted post-training on Russian enterprise traffic. The larger model still wins on long-context (ruMultiChallenge 46.2 vs 34.1), English (Arena Hard 2: 71.0 vs 60.8), and multi-step retrieval (SmartSearch 0.669 vs 0.557).
What reward hacking problems occur when training LLMs with GRPO on verifiable rewards?
This paper documents three: (1) semantic collapse — the model emits minimal, semantically empty completions that pass formal verifiers (e.g., answering an acrostic constraint with just the initials); fixed with a reward-model quality penalty relative to a teacher baseline. (2) Over-calling — a binary exact-match tool-call reward teaches the model to emit a call whenever in doubt; fixed by injecting synthetic 'no applicable tool' cases into the training mixture (irrelevance detection 78.3 -> 88.5). (3) Verbosity hacking — reward models prefer longer answers; fixed with a multiplicative length penalty against a per-prompt teacher baseline plus a raised KL coefficient (response length 2064 -> 1220 tokens with higher benchmark score).
What is SLERP model merging and why merge RL experts instead of joint training?
SLERP (spherical linear interpolation) interpolates two models' weights along the arc between them, preserving geometric structure better than naive averaging. The authors trained three GRPO experts (general chat, instruction following, function calling) from one shared SFT checkpoint and merged them in two stages ((IF+FC) first, then +general). Joint GRPO on all rewards caused cross-domain interference: at 32B it dropped ruBFCLv3 to 60.88 vs 65.96 for the merge and ruIFEval to 0.770 vs 0.799; even the best warm-started joint run needed 1.7x the training budget to approach parity.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- The provided text is a sampled excerpt with three '[...sampling gap...]' omissions (part of Section 5's SFT ablation, mid-Appendix D, and part of the references); judgments avoid relying on gap content and no claim of absence is based on it.
- Table 6's column header and value alignment is garbled in extraction; per-column assignments used in this review were cross-checked against running text and Tables 4, 18, 25, 28, and 29 (e.g., in-house Arena 69.57/65.83 and ruBFCLv3 65.96/64.42 are confirmed in Section 5 prose).
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.