Task validator filter

[p-ferreira]

Adds task validator filter to check if the proposed task is being accomplished.

This PR seeks to bring more consistency to the CoT developed by the validators and the miners, preventing tasks to be performed out of their prompts.

This is an initial simple keyword check verification model that can be evolved to more sophisticated use cases with real model implementation.

• adds --neuron.task_validator_off flag
• adds task validator model
• updates event structure to support task validator
• adds unit tests for task validator and updates event tests
• includes and complements prompt engineering work from #110

[mrseeker]

This can easily be bypassed by avoiding to mention "Question:". You might need to alter it to lowercase only and rethink the way questions are asked. The AI could also use "reason" or "thought" or something in those lines, or even avoid the words Question or Answer by giving an alternative word with a similar meaning.

[p-ferreira]

This can easily be bypassed by avoiding to mention "Question:". You might need to alter it to lowercase only and rethink the way questions are asked. The AI could also use "reason" or "thought" or something in those lines, or even avoid the words Question or Answer by giving an alternative word with a similar meaning.

Well noted, I implemented the verification to be case insensitive. I totally agree with you, as the network evolves the attacks will also evolve and we will need more sophisticated ways to validate if the completion is related to the task being proposed.

As of now, this simple verification should take care of good part of the current "topic drifts" that we see popping up in the network, we will follow this issue closely to what's being generated in the network