ZKLoRA
Zero-Knowledge Verification of LoRA training on SOTA Models in 1-2 Seconds
In 2024, zero knowledge verifiability for machine learning seemed impossible. The latency overhead was too high. We did a research report on it here.
But, it is 2025. The year of AGI. And, we have made it possible.
Today, we’re open sourcing a frontier research, ZKLoRA. A zero knowledge protocol that allows verification of LoRA fine-tuning of open source AI models, in 1-2 seconds.
And not only for toy models, but for current state of the art open source models like llama 3.3 etc. with tens or hundreds of billions of parameters.
Want to try it yourself? Here’s the code : https://github.com/bagel-org/ZKLoRA
Want to see the benchmarks and curious how it works? Read the full research paper here : https://github.com/bagel-org/ZKLoRA/blob/main/paper.pdf
We are Bagel, an AI and cryptography research lab using cryptography to enable the monetization of open-source AI.
We believe ZKLoRA kickstarts a new era for verifiable and monetizable model training across untrusted networks. More on how ZKLoRA fits into the entire Bagel infrastructure to come.
 | A guest post by
|