Yandex

Russian search engine giant Yandex on May 30 released Yambda, the world's largest open source dataset for music recommendation systems, containing 4.79 billion anonymized user interactions designed to help developers create smart music services. Yandex collected data from nearly 28 million monthly Yandex Music users over a ten-month period, specifically 4.79 billion user interactions with 9.39 million songs, and the dataset includes key listener feedback on whether a song is good or bad, with all interactions timestamped for increased accuracy. Ya...

Russian technology giant Yandex has launched an open-source large language model training tool, YaFSDP, which claims to be up to 26% faster than existing tools. According to reports, YaFSDP outperforms the traditional FSDP method in terms of training speed, especially for large models. In terms of pre-training LLM, YaFSDP is 20% faster and performs better under high memory pressure conditions. For example, YaFSDP can achieve 21% efficiency improvement for Llama 2 with 70 billion parameters, and 100% faster for Llama 2 with 70 billion parameters.