←
Yandex1,450 posts
Yandex avatar

Yandex

@yandexcom

Official English account for Yandex. We build AI and put it to work at scale across Search, Maps, Cloud and devices. Research, open source and product news.

◷ Joined October 2012▧ 267 Photos and videos
Yandex avatar
Yandex@yandexcom·Nov 12

Meet Do it!, Alice AI's personal agent, now in beta. It can make phone calls, navigate websites, and handle multistep tasks through specialized agents, all coordinated by a powerful AI harness. Here's how we built it and how it works ⬇️ medium.com/yandex/do-it-how-…

Yandex avatar
Yandex@yandexcom·6h

Our TabM neural network helped researchers in China classify electricity load patterns in Sichuan, from normal consumption to renewable-energy loads and sudden demand spikes. TabM is our open-source architecture for structured data, designed to improve prediction accuracy with lower computational costs. A team at State Grid Sichuan Electric Power used it in a system that achieved 98.5% accuracy on real-world grid data. Understanding these patterns can help grid operators forecast demand and identify potential overloads. Working with structured data in energy, logistics, or healthcare? Try TabM yourself ⬇️ github.com/yandex-research/t…

Yandex avatar
Yandex@yandexcom·Nov 17

It's #InternationalStudentsDay! Since 2007, Yandex School of Data Analysis has trained almost 2.5K specialists in AI, ML and data science. Our alumni shine at ICML and NeurIPS, launch startups and advance healthcare and wildlife conservation. Here's to tomorrow's luminaries!

Post attachmentPost attachmentPost attachmentPost attachment
Yandex avatar
Yandex@yandexcom·Nov 16

Every December, you put on that same old song. Can your recommendation system pick up on that pattern? Our generative recommendation technology is built for scale, reliability, and efficiency. ARGUS processes long histories of user interactions while minimizing GPU memory use.

Post attachmentPost attachmentPost attachmentPost attachment
Yandex avatar
Yandex@yandexcom·Nov 13

Amazon started with human assessments of 5.5K product pages across 10 categories — establishing a benchmark for what good looks like. With VLM output and OCR scores as features, a CatBoost model learned to predict expert ratings with 94% accuracy and a 0.95 F1 score. More details on Amazon's blog: aws.amazon.com/ru/blogs/mach…

Yandex avatar
Yandex@yandexcom·Nov 13

Our open-source ML library now helps @amazon recognize great product listings at scale. Creating a product page takes insight and creativity. Evaluating thousands of them takes VLMs, optical character recognition — and CatBoost.

Post attachment
Yandex avatar
Yandex@yandexcom·Nov 11

In benchmarks on Llama 2 and Llama 3, YaFSDP reduced training iteration times by up to 26% compared with PyTorch's FSDP. Combined with other memory optimizations, it helped speed up Yandex's own LLM pre-training by 45%. The library is available as open source for developers and researchers worldwide: github.com/yandex/YaFSDP

Yandex avatar
Yandex@yandexcom·Nov 11

YaFSDP is designed to make GPU memory management and communication more efficient when training transformer-based models. Like standard FSDP, it distributes weights and other data across GPUs instead of storing full copies on every device. Each model layer gets the necessary data when it needs it. Our engineers introduced reusable memory buffers and carefully coordinated calculations with data transfers. This allows GPUs to exchange information while continuing to work, reducing idle time and unnecessary memory operations.

Yandex avatar
Yandex@yandexcom·Nov 11

Our technology helps train LLMs up to 26% faster — and it's open source. Training a model takes a lot of computing power, but coordinating the work across multiple GPUs can leave up to 30% of their capacity underutilized. GPUs spend time waiting for one another instead of training. This is why we developed YaFSDP and subsequently released it as open source.

Post attachment
Yandex avatar
Yandex@yandexcom·Nov 9

TabDDPM is an open-source diffusion model that learns patterns from real tabular data and generates new, statistically similar records rather than simply copying existing ones. In the Alzheimer's study, the researchers worked with genetic and demographic data from just 542 people. Using TabDDPM, they generated additional training examples, expanding the training dataset without recruiting additional participants. arxiv.org/html/2209.15421

Yandex avatar
Yandex@yandexcom·Nov 9

Yandex technology helped researchers at @Harvard and @txst study Alzheimer's-related cognitive impairment. Using TabDDPM, developed by Yandex and HSE University, scientists improved their model's sensitivity from 68% to 77%. frontiersin.org/journals/bio… This is one of many examples of Yandex technology contributing to medical research worldwide. We recently shared how researchers have used CatBoost for tasks ranging from predicting complications after heart attacks to distinguishing Crohn's disease from intestinal tuberculosis.

Yandex avatar
Yandex@yandexcom·Nov 5

Yandex’s Yambda-5B is the primary dataset for DLRMv4, @MLCommons’ next-generation benchmark for training recommendation models. Developed by researchers at @AMD, @Meta, @nvidia, and ByteDance, it measures how efficiently modern recommenders are trained. mlcommons.org/2026/10/dlrmv4…

Yandex made Yambda-5B public to help researchers develop and evaluate recommendation models at scale. One of the world's largest open datasets of its kind, it contains 4.79 billion anonymized interactions from 1 million users across 9.39 million tracks from Yandex Music.

0:18
Yandex avatar
Yandex@yandexcom·Nov 4

Yandex made Yambda-5B public to help researchers develop and evaluate recommendation models at scale. One of the world's largest open datasets of its kind, it contains 4.79 billion anonymized interactions from 1 million users across 9.39 million tracks from Yandex Music.

0:18