AI & ML interests

None defined yet.

Recent Activity

thrifts-gem-4i  updated a Space 13 days ago
csylabs/README
thrifts-gem-4i  published a Space 13 days ago
csylabs/README
View all activity

Organization Card

csylabs · ООО ЛИИ

Open-weight Russian domain LLMs — education, sports, healthcare, law, regional languages.

ООО «Лаборатория инновационных инициатив» builds and releases open-weight domain-specialized large language models for Russian-language professional use cases. Repositories created — weights in active fine-tuning, public releases rolling out through Q2-Q3 2026.

Roadmap

Model Base Domain Status Public release
ЛИИ-Спорт-Gemma-4-31B-Preview Gemma 4 31B (Apache 2.0) Russian sports — federation regulations, training methodology, biomechanics, anti-doping, sports medicine 🟡 SFT in progress June 15, 2026
EduLLM-RU-27B v2 Gemma 4 31B (Apache 2.0) Russian K-12 + ВУЗ education (ФГОС-aligned) 🟡 corpus prep August 2026
EduLLM-Chuvash-27B v2 Gemma 4 31B (Apache 2.0) Chuvash language education, bilingual RU↔CV 🟡 corpus prep August 2026
ЛИИ-Школа-Gemma-4-31B Gemma 4 31B (Apache 2.0) Russian K-12 deep specialization (subject teaching, ОГЭ/ЕГЭ prep) ⚪ planned Sep-Oct 2026
ЛИИ-ВУЗ-Gemma-4-31B Gemma 4 31B (Apache 2.0) Russian university-level academic register, methodology research ⚪ planned Q4 2026
ЛИИ-Право-Gemma-4-31B Gemma 4 31B (Apache 2.0) Russian commercial law, contract analysis, procedural ⚪ planned Q4 2026 / Q1 2027
ЛИИ-Мед-Gemma-4-26B-A4B Gemma 4 26B-A4B (Apache 2.0) Russian clinical text (administrative scope only, not diagnostic) ⚪ planned Q1 2027
ЛИИ-Мобайл-Gemma-3n-4B Gemma 3n-4B (Apache 2.0) On-device consumer Russian use cases ⚪ planned Q4 2026

Why open-weight, why Russian-domain

  • Federation deployment under 152-ФЗ. Clients in Russian sports / education / healthcare / law cannot use Anthropic / Google / OpenAI APIs from Russian IPs. Domain-tuned open-weight models hosted on Russian sovereign infrastructure (Selectel) are the only viable path for federation-level adoption.
  • Domain SFT > general capability. Our benchmark research (ЛИИ-Спорт-Bench-RU, EduBench-RU) shows that base 30B open-weight models trail frontier closed-weight by 1.5-1.7 points on raw scores — but the gap is concentrated in domain register (ВУЗ, СШОР academic vocabulary) which closes 60-80% with proper domain SFT.
  • Open methodology + open benchmarks. Every model ships with a public benchmark that anyone can re-run for ~$5-30 USD. See csylabs-org/lii-sport-bench-ru and csylabs-org/edubench-ru on GitHub.

Open benchmarks (already public)

  • 🏆 ЛИИ-Спорт-Bench-RU — 655 questions × 35 sports, top-3 LLM-judge ensemble methodology, 7-model leaderboard
  • 📊 EduBench-RU — Russian K-12 + university education, 50 prompts × 5-dimension rubric, 30+ models scored
  • 📊 bench.csylabs.com — live multi-tab leaderboard portal (sport / education / clinical / legal — sport tab live; others rolling out Q3-Q4)

Methodology references

  • 📰 Habr (RU) — Seven LLMs on the Russian sports domain (May 18, 2026)
  • 📰 daniel.csylabs.com — build logs, federation-facing strategy posts (RU/EN)
  • 📰 arXiv preprints — coming Q3 with full bench v1.0 + first SFT releases

Contact


ООО «ЛИИ» (Лаборатория инновационных инициатив) · ИНН 2100031165 · Чебоксары, Чувашская Республика, Российская Федерация · РКН оператор ПДн рег. № 52-26-257975

models 0

None public yet

datasets 0

None public yet