只需8GB显存，就能跑通SFT、DPO、GRPO全流程，轻松理解DeepSeek-R1背后的推理训练奥秘。
Article URL: https://github.com/pochenai/nano-llm-posttraining Comments URL: https://news.ycombinator.com/item?id=49133851 Points: 6 # Comments: 0
单卡10GB显存就能跑满血DeepSeek-R1-671B，支持全系N卡/A卡甚至国产卡，MOE混合推理更是绝活。
Article URL: https://github.com/ztxz16/fastllm Comments URL: https://news.ycombinator.com/item?id=48728290 Points: 3 # Comments: 0
Hugging Face开源项目，目标复现DeepSeek-R1，值得关注的一手代码实现。
Article URL: https://github.com/huggingface/open-r1 Comments URL: https://news.ycombinator.com/item?id=48489917 Points: 239 # Comments: 18
探索用LoRA和NEFTune方法高效微调DeepSeek-R1-8B，降低资源消耗同时提升性能。
arXiv:2606.10392v1 Announce Type: new Abstract: Financial named-entity recognition (NER) is essential for translating unstructured financial reports a…
DeepSeek-R1的"啊哈时刻"背后，是真正的推理还是高级模仿？本文通过全面实证对比，解剖人类与LLM的数学推理差异。
arXiv:2606.07410v1 Announce Type: cross Abstract: The emergence of "Aha moments" in large language models, particularly DeepSeek-R1-0120, has raised t…