Coordinator and Validator
Last updated
Last updated
The task will be published to the Monte Carlo blockchain in the form of the Task Scheduling Token (TST) and then scheduled by the off-chain coordinator for the miner to run.
The Task Scheduling Token Standard is a token standard customized for task scheduling. The task scheduling information will be stored on the chain, and the parameters and the output results will be stored off-chain.
There is a reputation mechanism for miners, which serves as an important reference standard for scheduling.
When registering with Miners, they need to submit their benchmark data as the basis for scheduling. Subsequent scheduling and reward allocation will be centered around the benchmark for reputation rating.
New devices will default to a full reputation, ensuring fair scheduling. Each time a task takes longer than the benchmark, it will reduce reputation. The reputation will gradually recover through subsequent qualified tasks.
For those who need highly reliable results and don't have high requirements of timing, we can prioritize the scheduling of miners with good reputations. For those with high time requirements, they can increase the gas price to match higher computing miners.
We can obtain the same output with the same input by fixing the random seed, thereby verifying that the output result is computed by the model. This has certain requirements for the model, it needs to adopt standardized data preprocessing technology, model architecture and training process. The existing mainstream AI models such as LLaMA and Stable Diffusion can all ensure consistency.
Optimistically, we assume that the committed result is correct. There exists a validation period during which the validators can verify the results. If the validation is passed, the provided result will be valid and accepted.
We will cooperate with DePHY, continuously collect the metrics data of the device, and detect outliers for the validator to find the fake miner more accurately. In the future, we will also consider incorporating opML for on-chain result verification.
A computing task can be allocated to multiple miners to ensure the reliability of task scheduling.
The system will reward based on the miner's benchmark and actual execution time.
If the results are verified to be incorrect, there will be no reward. The staked tokens will also be slashed as a penalty.
If the time consumed is longer than others, or the time spent is more than the benchmark's commitment, it will lead to a reduction in the weight of the reward.
At the end of the validation period, rewards are automatically given to the miners.
