本帖最後由 marlin12 於 2023-4-6 11:03 編輯
hkepcu1 發表於 2023-4-4 08:42 
How ChatGPT Works: The Architectural Details You Need to Know
呢篇野只係講咗ChatGPT既NLM(natural language model),怎樣用transformer和self-attention既方法,去跟人類以語言溝通(了解和生成語句)。
Why is ChatGPT so good?
呢篇先至講出ChatGPT優勝之處:
(因為有太多專業用語,唔識點翻譯,所以直接把重點用英文列出來)
> GPT-3, released in 2020, is a whopping 175B parameter model pre-trained on a corpus of more than 300B tokens.
> Next, fine-tuning the LLMs with instruction datasets - require many human annotators (with expertise in diff areas) take a long time to built
> Finally, use RLHF(Reinforcement Learning with Human Feedback ) to build up a reward model (rates how good the original language model response), and then use the reward model to fine-tune ChatGPT using reinforcement learning. |