DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning
Громкая статья от китайцев про модели DeepSeek-R1-Zero и DeepSeek-R1. DeepSeek-R1-Zero обучена исключительно на RL без SFT и демонстрирует отличные способности к reasoning. Однако у неё есть проблемы: плохая читаемость предсказаний и language mixing (прям вот так - текст на двух языках). DeepSeek-R1 решает эти проблемы благодаря multi-stage training и использованию cold-start data перед RL и достигает результаты сравнимые с OpenAI-o1-1217.
Плюс авторы выложили обе модели и шесть дистиллированных в open-source.
Кстати, первый автор в прошлом выиграл много соревнований по ML - возможно это внесло свой вклад.
DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning
Громкая статья от китайцев про модели DeepSeek-R1-Zero и DeepSeek-R1. DeepSeek-R1-Zero обучена исключительно на RL без SFT и демонстрирует отличные способности к reasoning. Однако у неё есть проблемы: плохая читаемость предсказаний и language mixing (прям вот так - текст на двух языках). DeepSeek-R1 решает эти проблемы благодаря multi-stage training и использованию cold-start data перед RL и достигает результаты сравнимые с OpenAI-o1-1217.
Плюс авторы выложили обе модели и шесть дистиллированных в open-source.
Кстати, первый автор в прошлом выиграл много соревнований по ML - возможно это внесло свой вклад.
Right now the digital security needs of Russians and Ukrainians are very different, and they lead to very different caveats about how to mitigate the risks associated with using Telegram. For Ukrainians in Ukraine, whose physical safety is at risk because they are in a war zone, digital security is probably not their highest priority. They may value access to news and communication with their loved ones over making sure that all of their communications are encrypted in such a manner that they are indecipherable to Telegram, its employees, or governments with court orders. Although some channels have been removed, the curation process is considered opaque and insufficient by analysts. Update March 8, 2022: EFF has clarified that Channels and Groups are not fully encrypted, end-to-end, updated our post to link to Telegram’s FAQ for Cloud and Secret chats, updated to clarify that auto-delete is available for group and channel admins, and added some additional links. As a result, the pandemic saw many newcomers to Telegram, including prominent anti-vaccine activists who used the app's hands-off approach to share false information on shots, a study from the Institute for Strategic Dialogue shows. But Telegram says people want to keep their chat history when they get a new phone, and they like having a data backup that will sync their chats across multiple devices. And that is why they let people choose whether they want their messages to be encrypted or not. When not turned on, though, chats are stored on Telegram's services, which are scattered throughout the world. But it has "disclosed 0 bytes of user data to third parties, including governments," Telegram states on its website.
from no
