Our system operates by routing messages through both the censored and uncensored models. Each incoming message is first analyzed for its adherence to our content moderation guidelines. We employ a scoring system that evaluates the message across key dimensions such as sexuality, violence, and hate speech. This score allows us to determine the appropriate model for generating a response.
Every message is assigned a percentage score based on its content. If a message flags high on any of the moderation criteria, such as a high percentage for violence or hate speech, it’s routed to our uncensored model. This model can handle more complex and nuanced discussions that may not be suitable for the censored version. However, we ensure that every response is accompanied by a sanitized memory transcript of the ongoing conversation.
One of the challenges we face is the need to sanitize these transcripts effectively. Sanitization involves translating the original conversation into a format that retains the same meaning but adheres to the content moderation guidelines of the censored model. This process is essential, as it allows us to maintain the integrity of the conversation while ensuring that it remains within the bounds of what is acceptable.
To achieve this, we utilize a multithreaded approach that allows us to manage four different transcripts simultaneously. Each transcript represents a different thread of conversation, with variations in content moderation requirements. Our system synchronizes these transcripts, ensuring that they are coherent and aligned with the ongoing dialogue.
This means that when we sanitize a transcript, we not only translate the content but also ensure that the context is preserved. The goal is to convey the same essential message without violating any guidelines, thus fostering a healthier interaction environment.
At Partyhat.ai, our multithreaded memory transcript sanitation is a testament to our dedication to creating a safe, intelligent, and engaging platform. By effectively managing two language models and ensuring coherent, moderated conversations, we can provide our users with a valuable experience that respects their voices while upholding community standards. As we continue to innovate, we remain committed to refining our processes and enhancing our platform for all users.
