Как мы адаптировали LLM для русского языка

Как мы потратили 2 месяца на адаптацию Qwen3-0.6B для русского языка. Написали систему с нуля на основе 8 научных статей из arXiv. Исправили 6 критических багов (от NaN в fp16 до архитектурных проблем). Получили +35% training speed и +60% inference speed . В этой статье - честный рассказ о том, что не работает из коробки, какие грабли ждут в production, и как мы их обошли. Мы - это я и мой друг =)

https://habr.com/ru/articles/964510/

#nlp #llm #machinelearning #RussianNLP #tokenization #pytorch #deeplearning #ProductionML #mawo

🔖 The Top 5 Papers About MLOps You Should Know

5️⃣ Machine Learning Practices Outside Big Tech: How Resource Constraints Challenge Responsible Development By Aspen Hopkins, Serena Booth

#mlops #productionml #devops #data #datascience #readinglist #mlopsengineer

🔖 The Top 5 Papers About MLOps You Should Know (Part 2)

3️⃣ Machine Learning: The High-Interest Credit Card of Technical Debt by D. Sculley, Gary Holt, Daniel Golovin, Eugene Davydov,
Todd Phillips, Dietmar Ebner, Vinay Chaudhary, Michael Young

4️⃣ The ML Test Score: A Rubric for ML Production Readiness and Technical Debt Reduction By Eric Breck, Shanqing Cai, Eric Nielsen, Michael Salib, D. Sculley

#mlops #productionml #devops #data #datascience #readinglist #mlopsengineer

🔖 The Top 5 Papers About MLOps You Should Know (Part 1)

1️⃣ Operationalizing Machine Learning: An Interview Study By
Shreya Shankar, Rolando Garcia, Joseph M. Hellerstein, Aditya G. Parameswaran

2️⃣ Socio-Technical Anti-Patterns in Building ML-Enabled Software by Alina Mailach, Nortbert Siegmund

#mlops #productionml #devops #data #datascience #readinglist #mlopsengineer

📝 What kind of MLOps team are you? [Part 3/3]
#mlops #productionml #dataops #mlsystems

In early starts-ups & even at the Small/Med Size business, teams are often a combination of the different modes & that's totally fine!

You don't always need a specialized team!

💡What's important to recognize is to know this framework exists for organziational alignment, as well as to know when teams can be spun out.

📝 What kind of MLOps team are you? [Part2/3]
#mlops #productionml #dataops #mlsystems

🔍 Zeroing in on the ones that oftentimes constitute the ML Org or the Data org:

⛑ Enabling teams - Help the DS & Product folks get those models out the door using the internal plateforms & capabilities provided by the CST

⚙️ Complicated Subsystem team - Focused on maintaining & expanding the extremely technical solution they own

👷🏻‍♀️The Platform Team - Owns unified & integrated experience.

📝 What kind of MLOps team are you? [Part1/3]

🗺️ In the world of "team Topologies" there are 4 types of teams.

🌊 Stream-aligned teams (ST) ---------> Data science & Product (for example)

⛑ Enabling teams (ET) ---------> ML Engineering

⚙️ Complicated Subsystem team (CST) ---------> The Kubernetes Team, the GCP team, the Terraform team, the Redis team, etc

👷🏻‍♀️The Platform Team (PT) ---------> The ML Platform Team, The Data Platform Team, etc

#mlops #productionml #dataops #mlsystems

🧠 Everyone else: <LLM Experts, producing multi-modal Gen AI systems. >

🤓 Me: <Still troubleshooting that lambda function to calculate Euclidean distance of lat/long columns in Polars Dataframe for a sample project in Colab. > 😅

-----
#datascience #mlops #productionml #ai #mlengineer

The tools we have today are better than the ones we had before and this is especially true in the #mlops world. We have more options than ever before (cc: MAD Turck Landscape) but confusion is just as high as it ever was.

#mlops #productionml #mlengineering #oss #devtools #python

👉🏻 Online Inference =/= Streaming

We're all aware of this right? That they're not the same thing?

#mlops #mlengineering #datascience #dataengineering #productionml #mlsystems #systemdesign