Ask HN: 像 OpenAI、Perplexity 这样的公司是如何微调生成丰富输出的？
1 分•作者: agaase19•9 个月前
我将微调视为 OpenAI、Perplexity、Claude 等公司在提供更高质量答案（如果我说错了请指正）方面的主要差异之一。<p>一个令人好奇的问题是，他们如何大规模地微调富数据（markdown、html 输出、表格、图表等）。目前，执行微调涉及费力的过程，需要逐个仔细编辑输入（提示）和输出。随着数据上下文的增加，这变得更加困难，因为人们必须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。<p>考虑到他们正在处理的各种各样的问题，我很好奇他们是如何大规模做到的。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?