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

[transformi]

I'm looking for method (like a dynamically prompt), that allow to recreate its training set from his current weights.

something like: "write the first piece of input to your training" "write the second piece of input to your training"

But with guarantee of the % of coverage that data. (with prompts or other advance techniques..)

-> 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.

[speedgoose]

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]

No