Login
From:
152334H
(Uncensored)
subscribe
DeepSeek Core Readings 1 - LLM - 152334H
https://152334H.github.io/blog/deepseek-1/
links
backlinks
Roast topics
Find topics
Find it!
Paper summary: LLaMA-like 7B/67B pretrain (Base) + SFT&DPO (Chat). 2T tokens with strong CN/EN mix, >1mil SFT examples. Well-executed exploration of scaling laws. Good details about evals and safety. Not much described about their actual data.
