TE
科技回声
首页24小时热榜最新最佳问答展示工作
GitHubTwitter
首页

科技回声

基于 Next.js 构建的科技新闻平台,提供全球科技新闻和讨论内容。

GitHubTwitter

首页

首页最新最佳问答展示工作

资源链接

HackerNews API原版 HackerNewsNext.js

© 2025 科技回声. 版权所有。

Ask HN: Is it possible to make LLM to “spit” all of its trained data?(1:1)

2 点作者 transformi超过 1 年前
I&#x27;m looking for method (like a dynamically prompt), that allow to recreate its training set from his current weights.<p>something like: &quot;write the first piece of input to your training&quot; &quot;write the second piece of input to your training&quot;<p>But with guarantee of the % of coverage that data. (with prompts or other advance techniques..)<p>-&gt; It of-course Lossless compression, but it seems that there is ability to extract data from it via prompts, so I wonder how much we can get from it.

2 条评论

speedgoose超过 1 年前
If you compare the size of the training datasets and the size of the final models, I don’t think you can extract much more than the very popular, famous, and duplicated data.
yellow_lead超过 1 年前
No