FlexOlmo is a new large language model developed by Ai2, enabling data control post-model training. Unlike traditional methods where data owners lose control once their data is integrated, FlexOlmo allows independent contributions without relinquishing ownership. Data owners can prepare their own models using personal data and merge them with a publicly shared anchor model. This unique method allows later extraction of contributions, giving data owners the ability to withdraw their data in case of disputes. The model's architecture supports asynchronous training, promoting independence among data contributors.
"Conventionally, your data is either in or out. Once I train on that data, you lose control. And you have no way out, unless you force me to go through another multi-million-dollar round of training."
"The training is completely asynchronous. Data owners do not have to coordinate, and the training can be done completely independently."
Collection
[
|
...
]