请问HN：像OpenAI和Perplexity这样的公司是如何对丰富的输出进行微调的？
我认为微调是像OpenAI、Perplexity和Claude等公司在提供更高质量答案方面的主要区别之一（如果我错了，请纠正我）。<p>一个令人好奇的问题是，他们如何在大规模下对丰富的数据（如Markdown、HTML输出、表格、图形等）进行微调。目前，进行微调涉及到逐一仔细编辑输入（提示）和输出的繁琐过程。随着数据上下文的增加，这一过程变得更加困难，因为需要仔细检查输入数据，并提供正确的输出，包括格式、语法、用户界面等方面。<p>考虑到他们处理的问题种类如此广泛，我对他们如何在大规模下做到这一点感到惊讶。你有什么想法吗？
查看原文
I see fine tune as one of the major ways companies like OpenAI, Perplexity, Claude companies differ when it comes to provide higher quality of answers (correct me if I am wrong).<p>One curious question is how do they fine tune rich data (markdown, html outputs, tables, graphs etc) at scale. Currently, performing fine tuning involves the laborious process of carefully editing inputs (prompts) and outputs one by one. Becomes more difficult as the data context increases and one has to carefully examine the input data and provide the right output including things like formatting, grammar, UI etc.<p>Considering such a wide variety of questions they are processing, it amazes me how are they doing it at scale. Any thoughts?