Forgive my ignorance, but can this help trust untrusted training output?

[matbeedotcom]

Forgive my ignorance, but does this system help us trust untrusted training output?

As in, we can create distributed training systems where we can validate that the training output hasn't been poisoned? Or does this solve something else?

It seems like it would work with training, but it could also pertain only to inferencing.

[ppotash]

Hi, I'm not quite sure I follow the question. By training output, are you referring to the finalized weights after running a training procedure? With our repo, the main focus is to be able to run a forward pass that uses private, separate LoRA weights, and prove that the forward pass computation that involved the LoRA weights is correct.