This scheme is severely flawed in many many ways. Here is a short list.<p>1. The contest creator can trivially cheat the model creators and steal the pot. They simply have to overfit a model to the data and then submit that overfit model. It will trivially win and beat the other solutions.<p>2. The datasets are usually the expensive and difficult parts of these things. You don't want to publish it openly where anyone can steal it. Also, storing data on the blockchain is super expensive.<p>3. It's extremely expensive to run models on the EVM. Sure, you aren't training, but even that final evaluation will cost a bunch of money. Especially because the EVM doesn't support floating point. Have fun implementing and using exp.